visual processing of optic acceleration
TRANSCRIPT
Vision Res. Vol . 32, No. 12, pp. 2313-2329. t992Printed in Great Britain. All riehts reserved
0042-6989 92 $5.00 + 0.00Copyright O 1992 Pergamon Press Lrd
Visual Processing of Optic AccelerationPETER WERKHOVEN,*I HERMAN P. SNIPPE,* ALEXANDER TOET*Receiued 10 February 1992; in reuised form 15 May 1992
We present data on the human sensitivity to optic acceleration, i.e. temporal modulations of the speedand direction of moving objects. Modulation thresholds are measured as a function of modulationfrequency and speed for different periodical aelocity uector modulation functions using a localizedtarget. Evidence is presented that human detection of velocity vector modulations is zot directty basedon the acceleration signal (the temparal deriuatiue of the velocity yector modulation). Instead,modulation detection is accurately described by a two-stage model: a low-pass temporal filtertransformation of the true velocity vector modulation followed by a aariancJ detection stage. Afunctional description of the first stage is a second order low-pass temporal filter having acharacteristic time constant of 40 msec. In effect, the temporal low-pass filter ii an integration oflhevelocity vector modulation within a temporal window of 100-l40msec. A non-triviat link of thislow-pass filter stage to the temporal characteristics of standard motion detection mechanisms will bediscussed. Velocity vector modulations are detected in the second-stage, whenever the variance of thefiltered velocity vector exceeds a certain threshold variance in either t-he speed or direction dimension.The threshold standard deviations for this variance detection stage are estimated to be l7o/o for speedmodulations and 9oh for motion direction modulations.
Motion Acceleration Speedmodulation
INTRODUCTION
Man is capable of interacting successfully with complexdynamic environments. This abil ity is due primarily topowerful neural mechanisms that have evolved to pro-cess optical motion information (see Nakayama, l9g5bfor a survey). Therefore motion perception has beenstudied extensively. Psychophysical research has shownthat the human visual system contains highly sensitivemotion extraction mechanisms (DeBruyn & Orban,1988; McKee, l98l ; Werkhoven & Koender ink, l99l)that map spatiotemporal image structure into explicitmotion information (e.g. velocity and direction).
Motion perception has traditionally been studiedusing spatiotemporal invariant (uniform) motion stim-uli. Relatively few studies have aimed at the humansensitivity to the spatiotemporal structure of motionfields or uelocity uector modulations. Although previousstudies helped to define methods and stimuli, none ofthem allowed defrnitive statements concerning humansensitivity to acceleration or mechanisms for detectinghigher derivatives of motion (Regan, Kaufman & Lin-coln, 1986). This scarcity of studies is surprising, since innatural vision, optical motion on the retina is generailyvarying in both space and time even if environmentalobjects move at a constant speed and direction. Struc-tured motion fields are not just an inevitable burden for
tlnstitute of Perception TNo, Kampweg 5,3769 DE Soesterberg, TheNetherlands.
tPresent address: Utrecht Biophysics Research Institute (uBI), BuysBallot Laboratory, Utrecht University, princetonplein 5, 35g4 CCUtrecht, The Netherlands.
our visual system. In fact, it has been shown that thespatial structure (Koenderink, 1986) and temporal struc-ture (Arnspang, 1988) of optical motion fields are ofmajor importance to the visual agent and are closelyrelated to egomotion and 3D shape extraction.
The study presented here focuses on the humansensitivity to temporal velocity vector modulations, thatis, the abil ity to detect temporal variations in speed ordirection (called optic acceleration).
The paradigm
A fundamental and intriguing question to be answeredis: does the human visual system contain specific acceler-ation detectors? In other words, do human observersdirectly assess the optic acceleration of a moving object(the temporal deriuatiue of the velocity vector function)or do they indirectly infer optic acceleration from vari-ations in the perceived velocity along its trajectory (bysampling velocities at different times)?
This question strongly resembles a classic debate in thestudy of uniform motion perception: are human observ-ers able to directly sense optical motion, or do they infermotion indirectly lrom the variance in object positionover time? Nakayama and Tyler (1981) have answeredthe latter question using a target with a periodically(sinusoidally) modulated position in time. Theymeasured modulation threshold amplitudes as a func-tion of the frequency (inverse period) of the positionmodulation function. They argued that modulationthreshold amplitudes would be independent of themodulation frequency when motion was inferred fromthe variance in position. However, when motion was
23t3
2314 PETER WERKHOVEN er a/.
assessed directly (e.g. the temporal derivative of theposition modulation function) threshold modulationamplitudes were expected to decrease with increasingmodulation frequency.
For low modulation frequencies (<2Hz), Nakayamaand Tyler found strong experimental support for a directassessment of motion. Modulation thresholds did notshow an invariance when expressed in terms of dis-placement. For higher frequencies (>2Hz), Nakayamaand Tyler found deviations from the expected depen-dence of modulation thresholds on frequency, pre-sumably as a consequence of some finite temporalintegration of the motion signal in the human motionsystem.
To examine optic acceleration, we adopt this elegantparadigm used by Nakayama and Tyler substitutingvelocity modulations for position modulations. That is,we use a target with a velocity vector modulated in timearound a certain mean velocity vector and measurethreshold amplitudes for the detection of velocity vectormodulations as a function of the modulation frequency.We study velocity vector modulations both in the direc-tion of the velocitl, vector (speed modulation) andorthogonal to the velocity vector (direction modulation).Invariant modulation thresholds as a function of modu-lation frequency would indicate an indirecr detection ofmotion modulation or optic acceleration.
General stimulus considerations
The choice of an adequate stimulus to be used in astudy on motion modulation detection is not trivial. Itis important to design the modulation detection exper-iment such that detection cannot take place outside themotion system in other dimensions than speed or direc-tion. In the following we list a few considerations regard-ing some widely used stimuli in motion experiments.
Sine-v'at,e gratings. Sine-wave luminance gratings area powerful tool for studying l inear systems and also forstudying motion perception. However. the use of movingsine-wave gratings leads to several problems. First, localspeed and local temporal frequency are inherently con-founded. As a result, a speed modulation of a movingsine-wave grating might be detected outside the motionsystem as a local modulation of stimulus temporalfrequency. For example, a detector with a spatiotem-poral separable response function, thus not tuned tospeed at all, would be sufficient. Second, with a one-dimensional spatial pattern. such as a sine-wave grating,it is not at all obvious how one could study motiondirection modulations. Third. moving sine-wave gratingsallow extensive spatial integration by the motion detec-tion system. This property makes all spatially extendedmoving patterns especially unattractive to study spatiallylocal modulations in speed or direction. Fourth, spatiallyextended moving patterns inherently stimulate motiondetectors at a range of eccentricities. Thus, a study ofmotion sensitivity as a function of eccentricity cannot bespecific.
Random pixel arrays. Another visual stimulus oftenused in studies in motion perception is a random pixel
array or "Julesz pattern" (Julesz, l97l). An importantproperty of a Julesz pattern is that its power spectrumis flat. Therefore, a moving Julesz pattern with a modu-lated speed function would not yield the temporal fre-quency cue discussed above. However, human sensitivityto temporal modulations is l imited by the fl icker fusionfrequency. As a result of this cut-off lrequency fortemporal modulations. the sensed energy of a movingJulesz pattern decreases when speed increases. That is,when speed increases. an increasing proportion of thespectral components of the moving pattern wil l yieldtemporal frequencies beyond the fusion frequency, thusconceivably reducing the apparent contrast of the stimu-lus. Thus. speed modulation for Julesz patterns mayprovide the observer with an apparent contrast modu-lat ion as a cue.
Furthermore, Julesz patterns are spatially extended.Hence, they yield similar problems for the study ofmotion modulations as discussed above for sine-wavegrat ings.
Localized torgets. We have discussed a few extraneouscues associated with spatially extended stimuli. Many ofthese problems are circumvented when using stronglylocalized targets, such as dots. A moving dot allowsfor the study of local motion perception (restrictedspatial integration) and for the control of eccentricity ofpresentation. Furthermore. local temporal frequencymodulat ion is not a cue for mot ion modulat ion detec-t ion.
However. an increase in dot speed can yield anapparent spatial stimulus extent (if the visual systemintegrates the stimulus over a fixed window in time). andalso a decrease in apparent contrast (if the visual systemintegrates the stimulus over a fixed window in space). Toget some grip on the possible contributions of theseextraneous cues, we studied motion modulation sensi-tivity using moving dot targets and blob targets(spatially blurred dots).
METHODS
Method speed modulations
This section describes the method for our study on thehuman sensitivity to temporal modulations of motionspeed.
Stimulus specifcations. The stimulus consisted of amoving luminous dot (well above detection threshold) ofI mm dia. The dot projected on the screen of a CRT wasblurred by a sheet of diffusing material which was placeddirectly in front of the CRT screen. We estimate thestandard deviation of the resulting isotropic luminance"blob" at l.5cm, thus its full width at half maximum(FWHM) at about 3-4 cm. The dot moved horizontallyacross the screen at a variable (modulated) speed fromthe leftmost point to the rightmost point of a horizontaltrajectory across a distance do.This single left-to-rightmotion is called a sv;eep (the distance do is the sweep-length). When it reaches the right end on its trajectory,the dot returned to the far left position on the trajectoryand continued its motion (the next sweep). The time to
OPTIC ACCELERATION 2 3 1 5
f inish one sweep is called the sweep-time (lo). At aviewing distance d,., the average dot speed u0 was:
uo: t i t arctan(dof d,) . (1)
One motion stimulus presentation consisted of foursweeps. Thus, the presentation time was 4ro. The dotspeed was modulated in time yielding a non-uniformperiodic function r'. '(r) with modulation frequency ra.Speed modulation functions u,(r) were either periodic(symmetric) triangular functions A(t) or periodic blockfunctions l l(r) (see Fig. l). The amplitude du, of themodulation functions was varied but was always smallerthan the average speed uo, such that the dot speed wasalways positive. The phase of the periodic modulationfunction at the start of the stimulus presentation wasrandomized.
In addition to the moving dot, we also provided theobserver with a stationary fixation dot (a green LED),placed at a distance equal to the sweep-length d0 abovethe center of the horizontal trajectory, thus makingeccentricity of presentation (e ) about equal to the lengthof the trajectory of the moving dot. The sweep-time fora particular experiment was taken to be such that onesweep contained a few cycles of the speed modulation.Hence. for low temporal modulation lrequencies
2 n
t . l I pu tse6 ( t ) | |
I T0
t l - - Trianqular^(t) r v_-Z,I \Z
t i m e
t l , -
a ( t ) ( \ s i ne
. l \ , ,t i m e 2;r
1
r ( t )
0
4
r ( r )
00 t ime Z i r
FIGURE l. A sketch of some modulation functions: puise shaped
d(r) , t r iangular A(r) , s inusoidal O(t) , b lock shaped t I ( r ) and sample
function )11), as a function of time l. For this i l lustration. all functions
are normalized such that their mean value over time is 0.5. their
temporal wavelength is 2n, and the modulation amplitude is 100%.
except for function (l). The function (l) is the velocity modulation
function as used in the "sample and hold" paradigm of Burr et al.(1986) (for this i l lustration also with mean 0.5, and temporal wave-
examined, a longer-sweep time was required. The par-
ameters as set in the different modulation experiments
are specified in separate parameter tables in the result
section.In the main experiments, speed covaried with the
eccentricity of the moving dot. To study the effects of
eccentricity and speed independently, we ran two control
experiments. In one, we varied eccentricity but kept the
viewing distance constant. In the second. we varied speed
but kept eccentricity constant.Apparatus. The speed modulation functions were gen-
erated by manipulating the position of the beam of a HP
1321A high speed graphic display (P31 phosphor).
The beam produced a I mm dia luminous dot on the
screen (well above detection threshold). A Wavetek 185,
5 MHz, function generator produced a saw-tooth
horizontal position signal n(/), as a function of t ime /,
which was fed to the X-channel of the HP 1321A.
The horizontal position of the dot was l inear with this
signal. Hence, the dot moved from left to right across
the screen unti l the saw-tooth reached its maximum(finishing one sweep) at which point it returned(invisibly) to the far left and started to traverse the
screen again (the next sweep). The amplitude of x(r)(and thus the sweep-length) across the screen was con-
stant. For a constant period l" of this saw-tooth signal
in time. the dot crossed the screen at a constant speed.
determined by the temporal derivative of ,r(r), and thusproportional with the reciprocal period j of .r(l). The
reciprocal period I of .r(r) (and thus the dot speed) was
modulated in time by a periodic modulation function
u,(t) with (modulation) frequency ar (using a HP 33254
synthesizer function generator). The modulation func-
t ion u.( t ) was ei ther a t r iangular funct ion A(t) or a block
function l l(r) (see Fig. l). Speed modulation L'y(r) varied
around an average speed t'o with an amplitude du- (see
Fig. 1). This set-up allowed easy adjustment of the
average speed. amplitude and frequency of the speed
modulat ion funct ion.The importance of t'isual -fixation. Pilot experiments
showed that visual f ixation during modulation detection
experiments is crit ical. Observers reported to have no
dfficulties in detecting modulations when tracking the
moving dot for conditions where detection failed under
visual f ixation. Obviously, pursuit eye movements facil i-
tate modulation detection. It is well-known that the
pursuit system is quite slow (cut-off frequency at about
I Hz). For speed modulation frequencies higher than
I Ha observers could not follow the exact speed modu-
lation, but might track the dot at its average speed. The
actual speed modulation would then become apparent as
a displacement in the retinal coordinate frame. Thus,
allowing the observers to track the dot would provide
them with a displacement cue. resulting in modulation
detection outside the motion system. In order to elimin-
ate this cue, visual fixation is crucial.
Note that in much of the older l i terature (e'g. Hick,
1950), but also in more recent l i terature (Burr, Ross &
Morrone, 1986) no mention of the observer's f ixation
condition is made.vR 32 i12-F
length 2z).
2316 PETER WERKHOVEN er a/.
Procedural information. Speed modulation thresholdswere measured in a modulation detection experiment. Inone session, observers viewed 18 stimulus presentationsof a modulated speed function u(r) (with an averagespeed u0 and a modulation amplitude dr,,) and 18presentations of an unmodulated (uniform) speed func-tion (having a constant speed uo). The order of presen-tation for these 36 trials in a session was randomized, aswas the phase of the modulation function for the trialsthat contained the speed modulation. The tasks of theobservers was to indicate for each stimulus presentationwhether they perceived a modulated or an unmodulatedmotion in t ime.
Usually 4-5 sessions with different adequately chosenmodulation amplitudes were sufficient to determine thespeed modulation detection threshold by data interp-olation. We defined the speed modulation threshold W,as the relative modulation amplitude dt.f uo at thresholdperformance (yielding 80% correct answers). Measure-ments were performed binocularly with natural pupils ina darkened room. No feedback was provided in eitherexperiment.
In one of the control experiments we studied speeddiscriminarion using the present experimental set-up. Ina session for speed discrimination, observers vieweduniform speed functions with a constant speed that wasei ther higher (us* dt , , ) or lower (uo- dt , , ) than theaverage speed uo of the ensemble of presentations. Ob-servers indicated whether the perceived speed was highor low. Before a session started, the motion stimuli wereshown on request to build an internal representation ofthe high and low speeds. The procedure for determiningspeed discrimination thresholds was otherwise similar tothe procedure for the modulation detection experiment.
Subjects. Five subjects with normal or corrected-to-normal vision participated in the experiments. Threesubjects. HS, PW, and AT are authors of this paper andhad foreknowledge of the design, and are experiencedobservers in psychophysical experiments involving opticmotion. The results of these main subjects are presented.The general findings were confirmed by two naive sub-jects, working on an hourly fee. There was no obviouscorrelation between subject experience and thresholdvalues.
M e thod direction modulations
This section describes the method for experiments ondirection modulation detection.
Stimulus. Similar to speed modulation functions, thedirection modulation functions were generated by ma-nipulating the position of the beam of the HP 13ZlAhigh speed graphic display. However, for directionmodulation functions, both the horizontal and verticalposition of the dot were manipulated.
The time dependent horizontal position x(r) of the dot(the X-channel of the HP l32lA) was driven by a Hp3325A synthesizer function generator. This generatorproduced a saw-rooth signal x(r) with a period A thatdetermined the sweeptime r0 and an amplitude thatdetermined the sweep-length d0 For this direction
modulation experiment, the period and amplitude ofx(r) were constant during a stimulus presentation, re-sul t ing in a constant hor izontal speed r ' . ( t ) - u0: dol to.
The time dependent vertical position 1,(r) of the dot(the )'-channel of the HP l32lA) was driven by aWavetek 185 function generator. The y(l) signal deter-mined the direction modulation. The position functions1'(r) were periodic with frequency cr.r. The vertical speedfunction r', (r) was simply the temporal derivative of vert-ical position t '(t). Thus, the modulation frequency wasro. The average vertical speed was zero. The amplitude ofthe vertical speed modulation function r,,(r) is written asdr',. As a result. the speed of motion 1'(r) of, the dot was:
[tmr ' ( l ) - r ' n / l + - .
V l ' o(2)
For a small vertical speed r', (r) relative to the horizontalspeed uo [ r ' , ( l ) ( r 'o ] . the average speed was approx i -mately constant [ r ' ( t ) I r ,o] . The direct ion 0(r) of morionas a function of t ime r is approximately l inear in r ',.(r)when r ' , ( r ) ( r 'n:
o( t ) :u r . ,un [ !1 ' ) l -u . (3)L l ' ' I l 'o
The average motion direction 0o in all experiments washorizontal: 0o : 0. The amplitude of the direction modu-lation function 0(r) is de - arctan(dt:,, lu).
Tr iangular posi t ion funct ions . ) ' ( / ) : ,1( l ) resul ted inblock shaped direct ion modulat ion funct ions 0(t) :fI(r). Sinusoidal functions .r '( l) : O(r) resulted in sinu-soidal d i rect ion modulat ions 0(r) : Q(r) [ integrated y( l )functions], but shifted a phase i backwards in time.Final ly, b lock wave posi t ion funct ions y( t ) : l l ( r ) re-sulted in pulse shaped direction modulation functions0(t) : d (r) . An i l lustrat ion of these posi t ion modulat ionfunctions and resulting direction modulation functions isshown in Fig. l .
Procedure. The procedure was identical to the pro-cedure for speed modulation detection experiments.Observers indicated for each motion stimulus whetherthe moton was modulated (non-uniform) or not. Twomain observers participated in the direction modulationexperiments (HS and AT). Two naive subjects confirmedthe findings for the main subjects.
Modulation direction thresholds are defined as thedirection modulation amplitude d0 yielding thresholdperformance (80% correct answers).
SPEED MODULATION DETECTION
Speed modulation detection dependence on modulationfrequency
Results. Speed modulation detection thresholds I,Z,. asa function of modulation frequency a-t for two differentspeed modulation functions and different speeds u0 arepresented in Fig. 2. The parameter settings for differentaverage dot speeds u0 are listed in Table l. Since the datawere very similar for the three main observers (PW, HSand AT). we averaged modulation detection thresholds
.........ts---t-
-+-#
...,....1.r
+
1.7 deg/s5 deg/s1 5 deg/s1 deg/s
2.5 deg/s5 deg/s
OPTIC .\CCELERATION
, l 1 1 0
Modulat ion Frequency, . , ; (Hz)
FIGURE 2. Threshold speed modulation amplitudes Wu as a functionof speed modulation frequency ro. Thresholds W, are the relative speedmodulation amplitudes (dt;,,,t,n) that yield 80%o correct answers. The(very similar) data of three observers (HS. PW and AT) have beenaveraged. Triangular symbols indicate (symmetric) triangular speedmodulation functions A(t1.The different sizes of the symbols indicatedifferent average speeds ro as given in the figure. Note, that the closedtriangles indicate a special condition in which the diffusing screen wasremoved such that the target was a luminous dot. Square symbolsindicate thresholds for (symmetric) block speed modulation functionstl(t). The * symbois indicate results obtained in a separate speeddiscrimination experiment. Observers indicated whether a uniformmotion stimuius moved at a high veiocity (r:o * dr',) or at a low velocity( t 'n - r / t ' . ) . As in the modulat ion exper iments. Wu: dt , . l t ,o. For the lef tspeed discrimination threshold in the figure the presentation time ofeach speed interval was I sec. for the right threshold it was l25msec.To facil itate a comparison with the speed modulation thresholds. thetwo speed dist'rininatiorr thresholds are plotted at a horizontal positionthat equals hal f their inverse presentat ion t ime. Hal f the length of theshown error bar for each data point corresponds to the square-root
var iance of the binominal orobabi l i tv d ist r ibutron for that ooint .
W, for this presentation. The modulation thresholds ofFig. 2 are presented as (relative) speed modulationthresholds (speed modulation Weber fractionsW,.: du,f uo). Triangular symbols indicate the triangularspeed modulation function A(t) and square symbolsindicate the block modulation function II(r). Opensymbols indicate that the moving target was blob l ike.Solid symbols indicate dot targets.
Consider triangular modulation functions. For lowmodulation frequencies (ro ( 2Hz) speed modulationthresholds for very different conditions ( I deglsec dottargets at 0.25, 0.5 and lHz: 1.7 deg,,sec blob targets atI and 2Hz: and l5deglsec blob targers at 2Hz) areidentical within measurement error (approx. 32%). Thissuggests that speed modulation detection thresholds atlow frequencies are constant and independent of fre-quency, speed and target shape. However. speed modu-lation detection threshold values do depend on the shapeof the modulation function used. Thresholds for thetriangular speed modulation functions ,4 (r) are approxi-
mately a lactor 1.8 higher (averaged over speed, modu-lation frequencies and subjects) than the thresholds(17%) for the block modulation function l l(r).
At high modulation frequencies (tt > 2Hz) the speedmodulation thresholds in Fig. 2 rise strongly with in-creasing modulation frequency for both triangular andblock-shaped modulation functions.
Discussion: threshold inuariance at low modulationfre-quencies. The frequency independence of modulationthresholds for low modulation frequency strongly sup-ports the hypothesis that modulation detection is basedon the magnitude of the speed modulation signal. Themodulation magnitude is independent of modulationflrequency. A detection mechanism based on the differ-ence in maximum speeds of the speed modulation func-tion is indeed expected to yield constant thresholds,independent of modulation frequency.
The invariance of thresholds for low frequencies rulesout the hypothesis that speed modulation detection isbased on the magnitude of the optic acceleratio,? signal.The optic acceleration signal is the temporal deriuatiu^e ofthe speed modulation signal. Hence. its magnitude islinear with the modulation frequency. Therefore. a de-tection based on the acceleration magnitude is expectedto improve with increasing modulation frequency. Ahypothetical acceleration detector (requiring a constantacceleration threshold for detection) would yield a hy-perbolic (inverse l inear) decrease of speed modulationthreshold in Fig. 2, which is not supported by the data.
The low lrequency plateau in Fig. 2 rules out anotherhypothesis saying that observers base detection on thespatial excursions of the moving dot with respect to itsaverage path (i.e. the path of constant speed ro). Accord-ing to this hypothesis, the speed modulations are de-tected whenever the excursions exceed a certainexcursion threshold. The magnitude of the spatial excur-sion is the temporal integral of the speed signal and islinear with speed and with the period of temporalmodulation. Thus, the "excursion" hypothesis predictsthat speed modulation thresholds decrease with decreos -
ing modulation frequency. This prediction is inconsistentwith the finding that thresholds are constant for lowmodulation frequencies (see Fig. 2).
In conclusion, the threshold invariance at low modu-lation frequencies strongly support the view that humanspeed modulation detection is based on the speedsignal itself [the relative magnitude du,, 'L'o of the speedmodulation function u(r)], and not on the temporalintegral of r '(r) (position), or the temporal derivative ofu(r) (acceleration).
I 1oo
B B osto8 6 0o-c
l 4 0Eo=E 2 0q)q)
A 0
TABLE L Parameter settings for speed modulation thresholds of Fig. 2
SPeedWave form uo (deglsec)
Sweep-length
4 (cm)Sweep-time
ro (sec)Distance Eccentricity
4 (m) e (deg)
A ( t )A ( t )A ( t )A ( t )I1(r )
1 . 01 . 75.0
15 .02 .5
+ J
30303030
4 . 1 01 .685.0415 .05.04
6.00r0.203.40t . t 23.40
'Jas ( o0ll roJ u^roqs s€^\uortJunJ uorl?lnpotu oql Jo [?^rolur peads qJeo qcqlv\roJ a ,{cuanbe:J 1B splor{ssrr{l uorlJotap uorlBlnpougll,!\ sploqseJql uorl€u[urJJsrp Jo uosrJBdruoc e selzlrlrJ-€J slql 'Qdll: co sarcuenba:; lurodurot ..luelelrnba,,rleql te (syoqu,(s ,, *,,) Z
'8lJ ur u.roqs are sploqsarr{lSurpuodsorroJ aql .Jesu
0001 ro gZl : J uorl€rnp Ie^ -Jalur ue JoJ u^\oqs se^\ lueluuedxa uorl?unurJJsrp IeAJOI -ur-o1(l Jno ur snlnturls uorlolu r.uJoJrun qJ€E .snlnurlsuoltoru orlt Jo saurl uorteluosa-rd 1ua:eglp ]B sploqsorq]uorleunurJcsrp peads onu peJnseou e16 .sllnsaur
'0,tf ',tp:'1U :paeds e8erene eql ol olrleleJ pessard-x0 ore splor{soJr.ll uorleurrurrosrp poads ssploqseJqluollcelsp uorlulnpotu roJ sV .('ep - o,z) ,/v\ol ro (,np + 0:t)q8rq sem 1op Surnoru ,{pu_rogrun B 3o peeds penracred aq}rer.{Ioq^\ alsJlpul ot peq sre^rosqo .uorlcetep uorlslnpotuJoJ s? dn-los JBluorurJedxa aures oql Bursn lueuuadxouorleururJssrp peeds 3 poruroJrad an '6sploqseJql uorlBu-[urJJSrp rllrt\ uorlJolep uor]Blnporu arudruoc ol JopJouI 'suorlrpuoc
lelueurredxa cgrcods aqt uo tuopuedapuouo aJe sploqsonll uorleurrurJJsrp peeds .,{1a1eun1roJun'(tgOI 'eoychJ :ggOt .u€qro
ry u,(nrgaq) stuarur:adxauolututunrrslp poads ur punoJ sploqserql uell (fg ;orotJeJ e ,{11ucrd,(t) :aq8rq gontu ere uouratap uottDlnpoutpaads JoJ sploqsaJql ,:anelroq ,,{l8ursr.ldrns sdeq:a4
'suroilud lop-uopupr polel
-npou peads roJ pourutqo ( I66l) Irlpperg pue uep.rous,(q pauodar elep luocal qlr,{\ }uatuaer8e luellacxo ur ere(Z '31.{ ur sarunbs uado) suortcunJ uorlelnpou {Jolg eqlroJ Qtrttx) sepnlrldue ploqsorql aqI .uotto.Tttory
uottoututttJslp paads sa uottJatap uottDlnpow paadg
'uorlre{-raqe^(\
uorlelnpou peads ,{cuanbarg-^\ol eql Jo enle^ eql roJIBcrlrJJ lou sr rlnurls Orlt Jo (run:1cods ,{cuenbarS :o) a:n1-Jnrrs lerlsds lJexa aqt leql s,^Aoqs (uor1cun3 uorlulnpourreln8ueur eql roJ zp1 1 w) (Z 'Alg aes) sqolq perrnlq pugslop snourunl roJ sploqsorr.{l uortelnporu paeds Jo oruop-uodsotroc eq] '.lanaznog .suorlglnpou poads ,{cuanb-a{ q31q rog r€q leruozrroq e se pa^racrad aq tq8ru:peeds tuelsuoJ te ]op e .elduexe roC .paads qlr.ln a8ueqclrlSrru aJnlJnJts lerteds panracrad aqt .oslv .uor]re]ep
uorlelnpou paods JoJ enJ lszJluoJ tua:udde ue Burplar,(unrtcads :a,r,rod 1e_roduetorteds aql ur safueqc qll,\,\popunoJuoJ are suorlplnpou paeds lpql uortrnporlul oqlur pen8re a,16 'sla8.rDt top sa sta4nt qolq :uotssnry7
'JOl3a10pacuerJel e ur oJueu.lo3tad ploqsarql JoJ uorlJunJ lJolqeql Jo IBI{1 soturl L'l x tt aq ot ssq uortrunJrelnEueul eqt Jo (r{rdap riorrelnpou) epnllldrue aql 'eJueH 'apntlldrue etues or{l Jo ane,r reln8uerrl B Jolsql saurl OaJql ecuprJs^ e S3r.{ uorlsunJ uorlslnpotuenu^\ {lolq e lsql aloN 'uorlJelep >1eed e Jo peelsurluu8rs paads perollg egl Jo uortJalop numtoa e Bul-unssu fq (Z):o .(uorlrunJ uortelnpou reln8uurrl E uuqtsuorsJnJxe 'ploqsoJr1l Jpau ,a8ru1
le aurl oJotu qcnuspuods uorlJunJ uorlelnpour {colq B leqt alou) uortrel-ep lead uo uorleturuns {llgquqord Jo stJa.ga aqllunorce olur 8ur>1et (t) fq panloser aq uuc ,{cuede:rsrpluareddu slql l?ql ^\oqs e/n uorlcas lapory eql uI
'(suorlcunJ uorlelnpol'u {3olqpuu teln8uerrl roJ splor{seJqt lenba Burpler,() uorlJunJuortelnpotu aq] go (4eed ;o) apnlrlduu eqt uo ,tldurspes€g aq louuec uortJalep uortelnporu oqt leql slsa8Ens,{y8uo-r1s U
'uorlcunJ uortelnpotu porequ ssed-mo1aql uo sale:ado leql uortoalep Jo ed,(t eqt lnoqe uorleru-roJul sl€0^0r uorls^Jasqo SFII .suortcunJ uorl?lnpotu{colq puu reln8uerrt qloq roJ potceJeun ratlg leroduelaqt qEnorqt ssed lpn luu8rs uortelnpou paads parotlgoql 'sercuanbar; uorlelnpou 1(ol roC .uorlcunJ peedslecrsfqd aqt ̂\olloJ o1 qSnoua lseJ sr ratlU sn{t sorcuanber3uorl€lnpou esoql lE osn€Jaq '3urra11g
lerodruel ssed_zno1Euuaprsuoo ,{q poolsJepun eq rcuuo) uorlcunJ uorlel-npou aql Jo ruroJ eql uo sploqsorr.{t Jo oJuopuadep aqt'renenoq (zH t
'3'a) sarcuanbary uorl?lnpou .r\ol lV']serotur
Jo eq .{eu anssl slqt 3o ,(pn1s rer{UnJ V'uorlslnpour zH I aql JOJ uorleurrurJJsrp 3 r{Jns u:o;-rod01 alqe are.r ,(aq] see:aq^\ ,zH
E uer{l :aqErq sarcuenbet;le {ldep uollelnporu lua:uddu lenbo qll,r\ suorlJunJuorl8lnpou -ruln8uerrl puB {Jolq 10 utalpd 1u:odue1 aq1ateunurrJslp ot elqu 0q ot tou pelrodo-r foqJ .sro^Jesqo
Jno Jo slrodar anrlcedsortur pue letu:o;ur fq pa1:od-dns sr tdocrad aqt saleuruop luauodruoc ,{cuanbergIstuauepunJ 3ql 13ql 133J aql .suorlJunJ uorl?l-npou o^u ar{t Jo sorJuanber3 leluaruepunJ er{t Jo sepnl-rldrue eql Jo g'l -; orlur aqt qtr^\ luauaet8z elqeuosearur sr suorlounJ uorl€lnpou )tcolq pue reln8uur:1 3osploqserqt a^rielar aql roJ g.l Jo ortet srql .(tuosatd arescruorrusq ppo aqt fluo 'pesn suorlJunJ uortelnpou aql.;o ,trleruur,(s aqt Jo asneraq leql elou) Qldap uorlelnporu%0U le uoAO 'q8no:ql passed sr fcuanbo:y leluaur-BpunJ eqt ,{1uo zHZ < co sorJuonbarg roJ lprll aleurlseel(\ 'paur?lqo suorlJunJ ploqserql uortrelep uorlelnpouoql Jo sJrlsrralJsrur{c ssed-,,rro1 leluau?punJ aql Eul-JaprsuoJ fq pools:apun aq uuJ Burpug srql .serJuanba:g
uorlelnpotu r{31{ lV 'suorl3unJ uoltelnporu {Jolq aqlroJ ploqserql eqt ueqt ;aq8lr{ g' I Jo roloeJ e ore suorlcunJuorlBlnpotu paeds reynSuer:l roJ sploqserr.{1 ,,(cuanb-e{ uotlelnpotu uanr8 ? lV 'stutoJ- oi'tD.u :uo$sn)fle
'suorlcos
Sumroylo; Jql ur pup erar{ palueseJd Btep uo pas€qpe^rrop are a8els (uorsrcap) puoJes aqt pu€ aEets (.raltgleroduat) tsrg or{1 ,(;rcads leql sreteruered oql .euo a8els3o leu8rs uortelnpou parellg aqt uo posgq efiets uotsrJ-ep e sr a8uls puooes aql .uortJunJ uorlelnpou peadseql uo Surterado retlu lerodual ssed-mo1 ropro qtu eJo slsrsuoc a8uls lsrg eql 'lapotu uorlcatep uorlelnpotuo8ets-oznl aql uo aleJoqpla e^\ .uorlJOS
lapol/{ aqt uI'sapnlrldup ploqsarql
pesgeJcur 8urp1ar,{ suorlJunJ uorlelnpou,{cuenberg qElqgo ,{Eraue aql seJnpor uorlelnpou oql go Burrallg ssed-,rol lerodrual V
'uorlJung peads (1ecrs,(qd) anrl eql Jouorsren (pere11g ssed-,no1) porrnlq,(11e:odruet p uo paseqsr uorlcalop uortelnpou peeds uerunq teqt Surunssz ,(qpootsrapun oq uuo sqt,{llucrEoloueurouor{d .sarcuanbo-r;
q8rq te pecnpar sr uorl3unJ uorlulnpou eql Jo apnlru-8eu aqt ',(lluareddy ',(cuanberg uortelnpotu Bursearcurqtr,,rr f18uorls osrr e
'49 ur sploqsorql uorlslnpoul peedseql (zHZ < ro) sarcuanbarg uorlelnporu q8tq lj -sarc-uanbatt q7y,1 to Suuatlgf lotodutat ssod- Mol i uorssrcsl6J
'/'2 /, NE^OH)UEA\ dEIAd8ttz
OPTIC ACCELERATION
TABLE 2. Parameter settings for speed modulation thresholds of Fig. 3
2 3 1 9
Wave form
Speeduo (deg/sec)
Distance
4 . ( m )Eccentricity
e (deg)Sweep-length SweeP-time
do 1cm; to (sec)
A ( t )A ( t )A ( t )A ( t )A ( t )
0.260.632.55 .07 .5
25.010.42.602.00r .35
0.521 .265.00
10 .015 .0
Z J
233535
22222
Results (see " * " symbols in Fig. 2) show that speed
-liscrimination thresholds (6%) are indeed much lower
rhan speed modulation detection thresholds (l7oh fot
block wave modulation functions). The 60/o speed dis-
crimination thresholds were independent of presentation
time.Discussion. At the longest presentation time each
uniform speed was shown 1000 msec in the speed dis-
crimination experiment, yielding a 60lo threshold. It is
interesting to compare this 6% speed discriminationthreshold with the 17o/o speed modulation threshold for
block modulation functions at 0.5 Hz modulation fre-quency. For a block shaped modulation function at thisfrequency, the presentation time of each speed intervalof the block function was also 1000 msec. Thus,
although the different speeds in both experiments werepresented at equal (long) time intervals. the thresholdsare markedly different.
The high thresholds for modulation detection may bea consequence of a fundamental problem observers havern segmenting the modulated motion stimulus into highand low speed intervals when speed itself is the onlysegmentation cue, as originally proposed by Snowdenand Braddick (1991). However, we propose an al terna-tive explanation (as discussed in detail in the Model andGeneral Discussion sections): high thresholds for modu-lation detection may be caused by the uncertainty (of
observers) about the phase of the speed modulationfunction.
The cut-off frequency dependence on speed
Motiuation. The cut-offfrequency @.(t'o) is defined asthe modulation frequenc,v yielding threshold detectionperformance (80% correct answers) for a given averagespeed ro and modulation amplitude dr'-. The data inFig. 2 suggest that this cut-off frequency is a functionof speed. For example. for the lowest speed tested(1.7 deg/sec), modulat ion detect ion thresholds increasesomewhat faster for increasing modulation frequency(e.g. at 4 Hz) than the threshold for an average speed ofI 5 degTsec.
To study this issue further, we measured cut-offfrequencies for a wide range of average speeds uo. Tofacil i tate a comparison of our data with the cut-offfrequencies for random-dot patterns used in the exper-iment of Snowden and Braddick ( 199 I ), we removed thediffusing screen yielding a luminous dot as a target. Thespatial power spectra of dot targets and random-dotpatterns are comparable.
Results. We measured cut-off frequencies @,(uo) for awide range of average speeds us by measuring the
percentage of correct responses as a function of modu-
lation frequency at modulation depth dt),:100% for
the triangular speed modulation function A(t). The
speeds and corresponding parameter settings are listed in
Table 2. It should be noted that target speed was varied
by varying the viewing distance d, (although for two
conditions the sweep-length d0 was slightly adjusted).
The solid symbols in Fig. 3 are cut-off lrequencies for
dot targets and show a clear increase of cut-off frequency
with stimulus speed. Open symbols are cut-off frequen-
cies for blob targets and are extrapolated from the
thresholds for triangular modulation functions in Fig. 2
using a temporal low-pass filter that is justified and
specified in the Model section. Because these extrapo-
lated data for observer AT and HS were very similar' we
averaged them for this presentation.
Discussion. We fitted the dependence of the cut-off
frequency o.(uo) for dot targets (solid symbols) on speed
uo to a power function:
@. (uo ) oc u6 (4)
and est imated the power exponent a :0.3 - 0.35. In the
General Discussion section, we discuss this power law in
terms of well known properties of elementary motion
detectors.A comparison of the cut-off frequencies for blob and
dot targets shows that only for the highest speeds used
(uo > 7.5 degisec) the cut-off frequency becomes pattern-
--.-..----l---*-
HS (do t )
AT (dol)
HS.AT {blob)
. 1 1 1 0 1 0 0
Speed, r'6 (degls)
FIGURE 3. Upper cut-off modulation frequency or0 as a function of
speed ro. Cut-off frequency r-r.t. is defined as the modulation frequency
yielding threshold performance (80% correct answers) for a modu-
lation amplitude of 100%. Cut-off frequencies are measured for
triangular speed modulation function A(t) and two individual observ-
ers (HS and AT). Parameter settings are listed in Table 2' Solid
symbols indicate cut-off frequencies for a iuminous target dot. Open
symbols are cut-off frequencies for a blurred (blob shaped) target and
are extrapolated from the data of Fig. 2 using the temporal low-pass
filter described in the Model section. Because these extrapolated data
for observers AT and HS were very similar. we averaged them for this
presentatlon.
1 0 0N
3jo
3 1 0ctotJ-
o
o1
2320 PETER WERKHOVEN et a/.
dependent (see also Watson, Ahumada & Farrell 1986)'
This is consistent with introspective reports saying that,
for sti l l higher average dot speeds of the modulation, the
percept was a "string of beads". The "beads" pre-
sumably correspond to the places where the stimulus
comes to an instantaneous standsti l l , thus allowing a
significant luminance build-up over time in a small
spatial region.
Disentangling t'iev'ing distance and eccentricitt'
Moti,^ation. In this section we report on a control
experiment to test our claim that the high frequency
cut-off we find for our speed modulation thresholds
(see Fig. 3) is caused by low-pass temporal f i l tering. [n
Fig. 3 we showed that the temporal cut-off frequency
depends only weakly on stimulus speed (4-8 Hz for
speeds < 7 deglsec). Therefore, it is tempting to assume
a temporal frequency l imit for the speed modulation
detection system. However, to support this conclusion
we have to tackle the following problem.
In the above experiment, speed was varied by varying
the viewing distance (see Table 2), thus covarying the
eccentricity of presentation with stimulus speed. The
spatial grain size of the visual system increases approxi-
mately l inear with increasing eccentricity (Watson,
1987), such that the spatial resolution for the spatial
speed variations of our stimuli decreases as a function of
eccentricity. Therefore. one could claim that the (near)
invariance of the temporal frequency cut-off can also be
explained by a constant spatial frequency l imit with
respect to the grain size of the visual system at the
eccentricity of presentation of the motion stimulus.
To disentangle the effects of viewing distance and
eccentricity we measured the cut-off frequency at a fixed
viewing distance and stimulus speed, but at different
eccentricit ies.Results. We measured the cut-off frequency at a fixed
viewing distance (d, :2.40 m) and stimulus speed
(4 deg/sec) for different eccentricit ies of presentation
(e :0 .5 , 5 , l0 and l5 deg) . The sweep- length do is 42cm
and the sweep-time /o is 2.5 sec. Cut-off frequencies for
a triangular speed modulation function with 80% modu-
lation depth are presented in Table 3' Table 3 shows that
the cut-off frequencies are virtually identical at all eccen-
tricit ies.Discussion. A correct explanation for the approxi-
mately invariant cut-off frequencies is indeed in terms of
TABLE 3. Cut-off frequency a.r. obtained for the triangular
speed modulation function ,,1(l ), with relative amplitude
80% and mean velocitY uo : 4 deg/sec, for three observers at
four eccentricities t
Frcentricity Cut-off HS Cut-off AT Cut-off PW
e (deg) a" (Hz) a, (Hz) a, (Hz)
Speed, r '6 (deg/s)
FIGURE 4. The dependence of upper cut-off frequency (rr, on speed
r, , for a t r iangular speed modulat ion funct ion,4( l ) wi th { ixed 80%
speed moduiat ion ampl i tude. Eccentr ic i ty was. f i red at l0 deg' Indiv id-
ual data for two observers (AT and HS are plot ted).
a temporal high-frequency cut-off. and not in terms of
an eccentricity-scaled spatial resolution l imit. Of course
this temporal frequency l imit cln be described as a
spatial l imit in units that scale with stimulus speed.
However, because of the scal ing in human mot ion v is ion
of the spatial grain size with stimulus speed. we believe
that such a descr ipt ion is equivalent to our explanat ion
in terms of a temporal resolution l imit.
Disentangling riev'ing distanc'e and speed
Motiration. The cut-off frequencies in Table 3 at
constant speed but varying eccentricity are invariant.,
whereas the cut-off frequencies in Fig. 3 at covarying
stimulus speed and eccentricity do show a slight (though
systematic) variation. We hypothesized that this small
variation in cut-off frequencies depends on the stimulus
speed. We tested this hypothesis explicit ly by measuring
cut-off frequencies at fixed viewing distance and eccen-
tricity but different speeds.Results. We measured cut-off frequencies for a tri-
angular speed modulation function with a modulation
depth of 80% at a .f ixed l0 deg eccentricity and flxed
viewing distance (240 cm). but at different speeds
uo. Because sweep-t ime /0 was constant (1.25 sec). the
sweep-length 4 was directly proportional to the dot
speed.Cut-off modulation frequencies are shown in Fig. 4 as
a function of the average speed uo (at f ixed eccentricity!).
As expected, we find a dependence of cut-off modulation
frequency on speed.Discussion. We fitted the dependence of the cut-off
frequency on speed to a power function [see equation(a)1. The exponent e that f its the data of Fig. 3 best is
est imated to be a :0.25 for HS and u : 0.30 for AT.
The absolute values of the cut-off frequencies at a fixed
speed in Fig. 4 can be compared with Fig. 3. The cut-off
frequencies of this experiment (measured with 80%
modulation amplitudes) are roughly 0.8 times the cut-off
frequencies in Fig. 3 (measured for 100% modulation
amplitudes). This can be explained by the fact that the
filtered modulation signal is proportional to the modu-
lation amplitude dt, times an attenuation function (see
Model section). This filtered signal has to exceed a
certain internal threshold for detection to take place.
1 0N
3
>o
o
ll-
o
o
1
0 .55
l 0l 5
n.a.5 .35.2n.a.
6 .05 .55 .35. ' l
6 .85.44.95 .8
The data for observer HS at 0.5 and l5 deg eccentricity were
not avai iable.
+-{---1-t-
1 deg/s2d.glg1.7 deg/s
OPTIC ACCELERATION 2321
Modulation Fr"f,r"n.y, ,u (Hz)
FIGURE 5. Threshold direction modulation detection amplitudes d0
as a llnction of modulation frequency cu. Weber fractions Wo are
simply related to d0 by the expression: Wo--tand?. Data of two
observers (AT and HS) have been averaged. Solid circles are data
obtained with a sinusoidal motion direction modulation function Q(l);
Open squares indicate a block shaped motion direction modulation
function ll(l) [corresponding to a vertical position modulation ,4(r)] '
Thus a higher modulation amplitude yields higher
modulation frequencies.Note, that the cut-off frequencies at the lowest average
speeds in Fig. 3 were measured at much smaller eccen-
tricit ies than the l0 deg eccentricity for this experiment.Thus. it seems that eccentricity of presentation is of small
relevance to speed modulation detection, even for speeds
that barely exceed the motion detectio,n threshold at the
eccentr ic i ty of presentat ion.Finally, it is of interest to note that human modulation
detection sensitivity at the lowest speed tested(0.5 deglsec at l0 deg eccentricity) is excellent when
expressed in terms of the spatial excursion of the modu-
lated motion path from the average motion path at
average dot speed 1'0. These spatial excursions did not
exceed 0.9 arcmin. which is approximately the hyperacu-
ity threshold that was found at this eccentricity for static
stimuli with an explicit nearby spatial reference available(Westheimer. 1982)!
DIRECTION MODULATION DETECTION
In this section. we study velocity vector modulations
orthogonal to the average velocity vector (direction
modulations) resulting in curved trajectories. Instead of
measuring speed modulation thresholds. we measure
direction modulation thresholds. The precise generation
of the direction modulation functions is specified in the
Method section. as is the definit ion of direction modu-
lation thresholds. Otherwise. the procedure and organiz-
ation of these experiments are quite similar to those of
the speed modulation detection experiment described
above.
Direction modulation detection dependence on modulation
frequencyResults. In Fig. 5, we present direction modulation
thresholds as a function of modulation lrequency for
two different direction modulation functions' Circles
represent sinusoidal direction modulations O(l), squares
block shaped direction modulations l l(r). Since direc-
tion modulation thresholds were very similar across
the main observers, we presented the averaged
thresholds for observers AT and HS. The parameter
settings for different velocities and functions are listed in
Table 4.Direction modulation detection thesholds for an aver-
age speed L)o : I deg/sec and sinusoidal modulation func-
tions (small solid circles in Fig. 5) are approximately
invar iant (d0:10.2- l2.3deg) for the range of f re-
quencies tested (0.25-l Hz).At high flrequencies, thresholds rise strongly as a
function of frequency rr.r for both sinusoidal and block
shaped direction modulation functions. The average
ratio of threshold amplitudes for sinusoidal and block
shaped modulation functions at high frequencies
a 2 2 H z i s 1 . 2 6 .Discussion: threshold int'ariance at low' modulation fre -
quencies. The threshold invariance at low frequencies
indicates that human direction modulation detection is
based on the amplitude of the direction modulation
function and not on its temporal deriuatiue (directional
acceleration); direction modulation amplitude is inde-
pendent of modulation frequency, whereas the derivative
is l inear with modulation frequency. Therefore, a detec-
tion based on directional acceleration would yield in-
creasing thresholds for decreasing low modulation
frequencies. which rs not observed.
On the other hand, direction modulation detection
might be based on the magnitude of vertical dot position
(the temp oral integral of the direction modulation func-
tion). However, a constant vertical position threshold
(relative to the mean position) would yield decreasing
threshold direction modulation amplitudes for decreas-
ing modulation lrequencies in Fig. 5, which is also not
observed.At I Hz modulation frequency, the 6.6 deg threshold
direction amplitude at t 'o : 2 deg/sec is smaller than the
l0'2 deg threshold found at t 'o : I deglsec' Although
here we compare only two data points we speculate that
the asymptotic level is speed dependent. Increasing
direction modulation thresholds at low speeds are not
surprising given the relatively low sensitivity of the
human motion system at low velocities (DeBruyn &
Orban. 1988).
1 5
1 0
ltooc)
F o ). 0 )
E ! to v
E q
.9oo.=
TABLE 4. Parameter seltings for direction modulation thresholds in Figs 5 and 6
Wave form
Speedr,o (degisec)
Sweep-length Sweep-time
do (cm) to (sec)Distance EccentricitY
4 (cm) e (deg)
O( / )o(r )i l ( , )d ( r )
1 . 02.0t . 7t . 7
6.006.006.006.00
4 . 1 04 . 1 03.433.43
43433636
4222
2322
In conclusion, the observed asymptotic behavior ofdirection modulation thresholds as a function of modu-lation frequency supports the hypothesis that detectionis based on the amplitude of the direction modulationfunction. The observed absence of mechanisms tuned tovisual acceleration seems consistent with a study onmotion after effects (MAE) by Schwartz and Kaufman(1987), who reported that "there is no MAE specific toadaptation for changing directions as distinct fromsimple mot ion".
Discussion: waue forms, Because the fundamental fre-quency of sinusoidal functions has a smaller amplitude(a factor of f) than that of block functions (at a given
modulation amplitude). thresholds are expected to be afactor of | : 1.27 higher for sinusoidal than for blockfunctions at relatively high modulation frequencies. Theaverage rat io (1.26) of threshold ampl i tudes for s inu-soidal and block shaped modulation functions found forthe frequency range ).-2 Hz, is in good agreement withthe ratio, suggesting a relatively low cut-off frequency.The strong increase of direction modulation thresholdsfor both sinusoidal as for block functions, suggests atemporal frequency l imit for the direction detectionsystem of approx. 2Hz. We found a similar temporall imit for speed modulation detection.
Discussion : direction modulations us speed ntodulations.To compare the (absolute) direction modulation detec-tion thresholds d0 with the Weber fractions for speedmodulations. we use the following Weber fraction Wofordirection modulations Wo: d"^,,lt,o: tan d0. Now. bothWeber fractions Wa: du; f uo and W,: dr,f t^o are ele-gantly expressed in terms of velocity vector modulationsand can be compared.
For example, one can compute that the Weber frac-tion Wofor block shaped direction modulation detectionat I Hz and speed Do : 1.7 deg/sec is approx. 9%. Thisis about a factor 2 lo*-er than the Weber fractionW,: l7o/o for block shaped speed modulation detection.Furthermore, if one takes into account the effects ofprobabil ity summation (or variance detection of thevelocity vector modulation), very similar ratios (a factorof 2) for speed and direction modulation thresholds canbe shown to hold for other modulation functions used(see General Discussion section).
The fact that the human visual motion system is moresensitive to direction than to speed is a well-knownphenomenon in motion discrimination experiments,where Weber fractions for speed discrimination aretypically twice the Weber fractions for direction discrimi-nation (Nakayama, 1985b; DeBruyn & Orban, 1988).Thus,-although absolute Weber fractions for modu-lation detection are a factor of 3-4 higher than Weberfractions for motion discrimination experiments-theratio of Weber fractions for speed and direction modu-lations is very similar (about a factor of 2) in both typesof experiments.
The invariant ratio of sensitivity of speed and direc-tion reveals a fundamental characteristic of the humanvisual motion system. Let's consider a motion systemconsisting of an ensemble of Reichardt correlators that
have spatially rotational-invariant prefi lters (i.e. spatialinput filters that are not onentation selective) (see Gliin-der, 1990). For such a motion system, the higher sensi-tivity to direction than for speed is l ikely to be relatedto the fact that in the motion direction both spatial andtemporal prefi lters contribute to a broadening of theensemble response correlation peak, whereas in theorthogonal direction the width of the correlation peak isdetermined solely by the spatial prefi ltering.
Discussion: lov' -pass tentporal fltering at high.frequen-cies. We proposed a low-pass temporal f i l ter to explainthe cut-off frequency for speed modulation detection.The temporal l imit found for speed modulations (2 Hz)was similar to the cut-off frequency of approx. 2Hzobserved here for direction modulation detection. In theModel section, we model this temporal behavior ofdirection modulation detection analogous to the way wemodeled that of speed modulation detection: a temporalfi l tering of the direction modulation signal which corre-sponds to a temporal f i l tering of the speed signal r ', (r)or thogonal to the mean veloci ty vector [see equat ion (3)] .
Spatial cues at t,ery high modulation frequenciesMotication. For modulation frequencies that far ex-
ceed the temporal l imit (2Hz) for detection of modu-lation by the motion system, thresholds are determinedby other cues outside the motion system. such as spatialcues. Although these thresholds do not reveal thecharacteristics of the motion system, they are of interest.
Results. We measured direction modulationthresholds at 8 and 100 Hz modulation frequency for aL7 deg/sec speed frequency. Thresholds were foundto level off for this frequency range, suggesting thatobservers made use of weaker spatial cues to detectmodulations, which were independent of modulationfrequency.
The threshold for the remaining spatial excursion cueat high frequencies is a measure of the dynamic spatialacuity orthogonal to the motion direction of a movingdot. We found it to be about 8 arcmin for our observers,which is a few times the 2-3 arcmin acuity limit for staticstimuli (Wertheim, 1894) at the eccentricity used in thisexperiment (3-4 deg).
Pulse -shaped direction modulation functionsMotittation. Here we make a short study of pulse-
shaped direction modulation functions d(r), yieldingsquare wave vertical position modulation functions II(r)(see Fig. l). Detection of pulse-shaped direction modu-lation functions is interesting because the correspondingvertical velocities are too high to be sensed by the motionsystem. Hence, modulation detection must be based oncues outside the motion system. Therefore, the detectionof pulse-shaped modulation functions is l ikely to makeuse of a spatial cue: the spatial excursion from the meanvertical position.
Filtered pulse-shaped direction modulation functionshave constant amplitudes for low modulation frequen-cies (when the response functions in the time domain forconsecutive pulses are well-separated), yielding constant
PETER WERKHOVEN er a/.
OPTIC ACCELERATION : ) z )
spatial excursions from the mean vertical position.
Therefore, if the spatial excursion is the cue for modu-
lation detection, thresholds are expected to be indepen-
dent of modulation frequency for low frequencies. We
tested this prediction by measuring modulation detection
thresholds (expressed as spatial excursions) as a function
of modulation frequency for pulse-shaped modulation
funct ions d (r) .
Results. The parameter settings for this pulse detection
experiment are l isted in Table 4.
Figure 6 shows that the spatial excursions yielding
threshold performance are approximately invariant
(2.3-2.8arcmin) for the range of modulation fre-
quencies tested (l_4Hz). Similarity in performance
for subjects AT and HS allowed averaging over these
observers.Discttssion. The finding that the spatial thresholds at
the I Hz modulation frequency are slightly higher than
those for 2 Hz may be explained by probability sum-
mation or by the nature of variance detection. At the
fixed stimulus presentation time used there are twice as
many "events" (pulses) at a 2 Hz than at a I Hz modu-
lation frequency. Consequently, probabil ity summation
or variance computation can take place across more
events yielding lower thresholds. Obviously, there is a
limit to summation such that no improvement occurs at
even higher modulation frequencies.Note that the spatial modulation thresholds (spatial
excursions\ (2-2.5 arcmin) are considerably lower than
the dynamic acuitt ' (8 arcmin) determined earlier in
this paper at the same speed (1.7deg1sec). However,
they are sti l l higher than the hyperacuity thresholds
(0.2-{).5 arcmin) that have been measured for stationary
spatial configuration at similar eccentricit ies (West-
heimer. 1982). A more elaborate experiment that was
performed on pulse shaped speed modulation functions
will be presented elsewhere.The fundamental frequency components of block
shaped and triangular shaped position modulation func-
tions (at a given modulation amplitude) have relative
amplitudes i. With a low cut-off frequenc.v of I Hz. these
fundamental frequencies are expected to dominate the
detection thresholds at high modulation frequencies.
Therefore, we expect relative thresholds of I ry 0.64. This
can be verified by comparing thresholds for block
shaped position functions (see pulse shaped direction
modulations in Fig. 6) with the triangular shaped pos-
ition functions (see block shaped direction modulations
in Fig. 5) at, for example. 4 Hz modulation frequency'
This average ratio across observers is 0.67 supporting
our claim that we deal with positional cues and a
low cut-off frequency such that the fundamental fre-
quencies dominate detection of high lrequency modu-
lation functions.
MODEL
we present a model for the detection of velocity vector
modulations in the human visual system. The model
consists of two stages. The first stage is an nth order
low-pass temporal filter that operates on the velocity
vector modulation function. This fi l ter is characterized
by its order n and a characteristic t ime constant z. The
second stage is a decision stage based on the filtered
modulation function.In this section, we wil l show that our data provide
strong experimental evidence that the decision stage is a
uariance detection stage. The single parameter that
specifies the variance detection stage is a variance
threshold and can be estimated from the data. Further-
more, having knowledge about the decision stage, we
can estimate the parameters that characterize the first
(low-pass temporal f i l ter) stage.
The decision stage: cariance detection
The amplitude detection thresholds presented in this
paper have been interpolated based on detection proba-
bil it ies as measured using a method of constant stimuli.
Detection thresholds are the modulation amplitudes at
threshold performance (80% correct answers), and thus
form one parameter to characterize the full psychometric
functions available. However, the shape of the psycho-
metric functions reveal the parameter used in the
decision stage (e.g. velocity, squared velocity, etc').
Therefore. it is of interest to examine the shape of the
psychometric functions.Due to the noise, associated with the stochastic (bino-
mial) nature of the observer decision process. we need a
large number of elementary decisions for each modu-
lation amplitude in order to discriminate small shape
differences of different psychometric functions. There-
fore. we averaged psychometric functions, for all par-
ameter settings used in the modulation detection
experiments described above, as a function of normali:ed
velocity vector modulation amplitudes i. With a normal-
ized vector modulation amplitude (, we mean a modu-
lation amplitude for a given parameter setting, divided
by the modulation amplitude yielding threshold per-
formance for that part icular set t ing: C:@tlu1JlW' 'Assuming that velocity vector modulation detection
is ruled by a single detection process' psychometric
goooL J
E
F ^. Y Co ' = t= 6( ' ' ; ) L
x G l
( ! I'-
a
1 2 4
Modulat ion FrequencY, * (Hz)
FIGURE 6. Modulation thresholds for the detection of pulse shaped
direction modulation functions d (l ), expressed as vertical spatial
excursiotts of the block shaped vertical position modulation fl(l ).
Spatial thresholds are shown as a function of modulation frequency cr.t.
The average speed t'0 was l.Tdeglsec and the eccentricity of presen-
tation was 3.4deg. Data for two observers (AT and HS) have been
averaged.
PETER WERKHOVEN er a/.2324
functions for all parameter settings will be identical ifplotted as a function of normalized amplitudes!
The resulting psychometric curve that describes all
experiments reported in this paper is shown in Fig. 7.
The horizontal axis in Fig. 7 represents normalized
modulation amplitudes (. The ordinate represents thepercentages correct for a small range of normalizedmodulation amplitudes clustered around a range ofplotted normalized amplitudes of the data points. Halfthe length of the shown error bar for a normalizedamplitude i corresponds to the sqare-root variance (or)
of the binomial probabil ity distribution for that point:
o, :1t .11 -p51., ,6, , wi th p, the ordinate for modu-lation i. and ni the number of elementary observerdecisions. Each individual data point is based on about700 elementary (yes/no) observer decisions (n,: 700).
We find an excellent fit of this psychometric functionusing a standard error function or (scaled) standardnormal distributi on Erf(z):
r f',':
Erf( : ) : + | , - t iz 'zt 4 i , (5). \ / L t L J
- q
that has the square normalized modulation amplitude(: : (2) as i ts argument (Xt :5.2 wi th 7 d. f . ) . The scalefactor ̂ i rs constant: ̂ l :0.84, and causes Erf(z) to be80% for z : | (a constraint, set by our threshold defi-nit ion). Note that the function Erf((2) used for the fit hasno free parameters! The evidence for this particularshape of psychometric function is very strong since a fitwith functions Erf(O or Erf((t) yields unacceptable
chi-square values of X' : 152 and 7?: 55 respectively.Thus, assuming that a standard error function is a validdescription of the psychometric function associated withthe final observer decision process (Green & Swets,1966), the empirical function shown in Fig. 7 stronglysuggests that modulation detection is based on thesquare modulation amplitude. i.e. the uariance of the(temporally f i l tered) velocity vector modulation.
A variance detection process is certainly not exclusivefor velocity vector modulation detection and has. forexample. also served to explain human sensitivity totemporal fluctuations in the luminance domain (Rash-bass. 1970: Koenderink & van Doorn. 1978). The choiceof variance detection for these visual tasks is not surpris-ing since variance detection has been shown to beoptimal for a statistical point of view in a number ofinstances in which the visual system is uncertain aboutsome aspects of the stimulus to be presented (Green &Swets. 1966), e.g. the phase of per iodical modulat ionfunctions. This uncertainty may force the observer to usethe autocorrelation (variance) of the velocity modulationsignal. instead of the more efficient cross correlation ofthe modulation signal received with the signal expected(Green & Swets. 1966; Burgess & Ghandeharian. 1984).
We argued that the uncertainty of the observer aboutthe phase of the modulation function accounts forvairance detection and thus for our finding that speedmodulation detection thresholds (l7o/o) are much higherthan speed discrimination thresholds (6%). However. analternative explanation for variance detection is in termsof the diff iculty of the visual system in segmentingmodulated motion paths in temporal segments of differ-ent velocities in the absence of cues other than in themot ion dimensions (Snowden & Braddick. l99l) . Todecide on this issue, experiments are useful in which theobserver is provided with explicit cues that define thephase of the motion modulation signal andlor allow fora segmentation of the stimulus.
Interestingly, such experiments have been performedin another domain. In stereo vision (disparity process-ing), for example, the absence of an explicit segmenta-tion cue (i.e. defined outside the stereo domain) can leadto dramatic increases in the disparity discriminarionthreshold (McKee, 1983; Fahle & Westheimer, 1988). Inthe domain of two-dimensional luminance pattern per-ception. on the other hand, an explicit segmentation cuehas been observed to force the visual system in aprocessing mode yielding higher discriminationthresholds than obtained without segmentation cues(Watt , 1985)!
Rashbass (1976) has shown that a uariance detectionprocess based on the square modulation amplitude yieldsidentical detection performance to an alternative detec-tion model consisting of peak detection based on thelinear modulation amplitude, with a detection prob-ability summation that is governed by a psychometricfunction similar in form to the empirical curve shown inFig. 7.
We have shown that modulation detection is accu-rately described by a variance detection. based on a
1 0 0
90
Ec 8 0oo
3 7 0o-
l n
0 . 6 0 . 8 1 . 0 1 . 2 ' 1 . 4
Norma l i zed Amp l i t ude , €
FIGURE 7. Probabilit ies P of correct answers as a function of thenormalized velocity vector modulation amplitude (. With a normalizedvector moduiation amplituded i, we mean a modulation amplitude fora glven parameter setting divided by the modulation amplitudeyielding threshold performance for that particular setting. The datapoints are collected using all speed/direction modulation detectionexperiments described in this paper. The horizontal axis representsnormalized modulation amplitudes (. The ordinate represents thepercentages correct P for a small range of normalized modulationamplitudes clustered around a range of plotted normalized amplitudes
{, of the data points (solid circles). Half the length of the plotted errorbar for a normalized amplitude (, corresponds to the square-rootvariance or standard deviation (o,) for that point (see text). Becausewe normalized the amplitudes to the threshold amplitude, the curve isexpected to reach threshold (807o correct answers) for i : I and reachchance ievel (50%) for { :0. The solid curve is a best fit of thepsychometric function Erf(z) to the data, taking the squared modu-lation amplitude z : 4, as its argument. The dotted and dashed curvesare the expected psychometric curves for arguments ; : ( and z : g I
respectivelv. and fit less well.
OPTIC ACCELERATION
low-pass transformation of the velocity vector function,
that is. velocity vector modulations are detected
whenever the variance (after filtering) exceeds a certain
internal threshold. The internal thresholds are most
easily derived from threshold modulation amplitudes for
block shaped modulation functions at a low frequency,
when the modulation function is nearly unaffected by the
low-pass fi l ter. This is because the variance of block
shaped functions is equal to the square modulation
amplitude. Therefore. the internal threshold for variance
detection is equal to the square of the threshold modu-
lation amplitude measured for these functions. For
example. the minimum standard deviation (square-root
variance) yielding modulation detection is W,: 17o/o for
speed modulations (see Fig. 2) and W :9o/o for direc-
tion modulations (see Fig. 5)'
The filter stage: a second-order lov'-pass .filter
The data presented in this paper suggested that the
detection of velocity vector modulat.ion functions is
based on a temporal low-pass fi l tered version of the true
(physical) stimulus velocity vector function. The effects
of such a low-pass temporal f i l ter on detection perform-
ance depends on the type of decision stage that follows
temporal f i l tering. Now we have specific knowledge
about this decision stage (see above), we can estimate the
temporal characteristics of the first (f i l ter) stage'
We model the temporal low-pass fi l ter as a standard
temporal n th order low-pass fi l ter and estimate its order
and its characteristic t ime constant from the dependen-
cies of modulation thresholds on modulation tiequency.
Methocl. A standard temporal nth order low-pass fi l ter
has a pulse-response function:
8 ( r ) : - f * ( l ) ' e - ' ' ( n > o ) . ( 6 )r l n - l ) ' \ r /
and a transfer function A(c,r ) of modulation frequency 1a:
S@t) : [ + (2non) : ]
This f i l ter reduces the ampl i tude of the modulat ion
functions. For example. a sinusoidal modulation func-
tion with frequency o and amplitude .4 passing the l i l ter
g(r) wi l l have a reduced ampl i tude A!(ot) . The detect ion
of such a modulation signal takes place in a variance
detection stage as described above. At threshold. the
variance of this fi l tered sinusoidal signal is equal to a
th resho ld var iance o i : A ' 'E t (o t )12 : oa ( reca l l tha t the
var iance of a s inusoidal s ignal equals hal f i ts square
ampl i tude).This threshold oi can be estimated from an empirical
modulation detection threshold w(an) at a low modu-
lation frequency @o ( l lQnt) for which g(r'to) = l ' For
example. for block shaped modulation functions the
variance of the function at threshold amplitude W(algo)
is exactly wt(') ' Thus' oi: w'(con) for block shaped
modulation functions.To estimate the time constant r, we consider the
threshold amplitudes at modulation frequencies r.r.r, and
co, for which the higher order spectral components of
the rno,Julation function can be ignored- In those cases,
2325
the modulation functions are approximated by their
fundamental (sinusoidal) components. For example, for
block shaped functions with (threshold) amplitude
Wko,), the amplitude of the |undamental sinusoidal
component is f W(a,). At threshold, the variance of the
filtered fundamental equals the detection threshold oi:
: o i . (8)
For frequencies |i. i) (2m )-
"
we could use the asymp-
totic behavior E@) = (Zna,t)- ' to estimate r and n
analytically from two data points at high frequencies.
However, we can not use this approximation a priori and
use a numerical approach. First, we rewrite equation (8)
to solve r? as a function of n and crr,:
I [ / . , ,Ew,1 . , ) \ t ' _ , - l:,^;iL(H) -rl (e)
The t ime constant t is constant. Thus. r(co, , n) is ex-
pected to be the same for any two different (sufficiently
irigft) modulation frequencies cDr and crr' ' Hence, the
order n of the fi l ter is the solution of the equation:
( l 0 )
Il| ""t', l] t | + (2nw'r')l -'
t ( ( t r . n ) : t ( o t r , n ) .
W e c h o s e t h e l o w e s t n f o r w h i c h l r ( o ' , n ) - t { a t , n ) l i
[ r ( ro , , n ) + r (@t ,n) ] i s <20o/o . The t ime cons tan t t i s
iaken to be the average of the two values r(@r,n) and
r ( r o 2 , n ) a t t h i s n .
Estimation oJ'Jilter parameters for speecl modulation
thresholcls. Consider the speed modulation thresholds tor
block shaped functions in Fig. 2. We use the threshold
at low modulation frequency @o: I Hz for the esti-
mat ion o f oo "
o0 : W, ( l \ : 17o /o ' Fur thermore ' we use
the two thresholds at high modulation frequencies
()) t : 4Hzand cr.r ' : 8 Hz with thresholCs W,(a,) :30o/o
and W,(ro,) : 8l%'
Using the method described above, we find that a
value n :2 and a t ime constant r : 33 msec adequately
model the dependence of speed modulation detection
thresholds w,(a) as a funct ion of modulat ion f requency
@. In fact . we found r(cor.2) : 30'6 msec and
r ( @ : . , 2 ) : 3 6 . 1 m s e c .With r :33 msec. th is second order low-pass f i l ter
corresponds to a value of approx. 90 msec for the full
width at half maximum (FWHM) of the pulse-response
of the speed integration fi l ter yielding an integration
(smoothing) of the physical speed signal in the human
visual system within roughly a 100-140 msec temporal
window.Estintation of ftter parameters for direction modulation
thresholds. Similar to the previous section. we estimated
the order n and time constant r from the direction
modulation thresholds for block shaped functions as
presented in Fig. 5. We used the threshold^Wo(a) - -8 '7o /o
fo r (D0: lHz ' and Wakt t ' ) : l l ' 4o /o
(ar ' : 2Hz) and W6(ror) : 24'9% (ar:4Hz) '
Using the method described above' we find that a
value n :2 and a time constant r : 42 msec adequately
model the dependence of direction modulation detection
threshold s w6(a) as a function of modulation frequency
2326
()). In fact, wer( :@2,2) : 49 msec.
This time constant for the low-pass filtering of the
direction modulation signal is only slightly higher than
the time constant (r :33 msec) estimated from the
low-pass filtering of the speed modulation signal. How-
ever. as we showed in Fig. 3, this small discrepancy may
be a consequence of the different speed ranges used for
the determination of the temporal filter characteristicsfor direction and speed modulations.
The overall similarity of the characteristics of human
detection of speed and direction modulations and the(near) equality of the integration time-constants derived
strongly suggests a detection system that monitors the
full (temporally filtered) velocity uector.
GENERAL DISCUSSION
Euidence .for indirect optic acceleration detection
We presented a study of human sensitivity to optic
acceleration and have been unable to find any evidence
for a visual mechanism that directl l 'detects optic accel-
eration. i.e. the temporal derivative of the velocity vector
modulations. Instead we find strong evidence that modu-
lation detection is based on the amplitude or modulation
depth of a temporally filtered velocity vector modulation
signal. The temporal characteristics of the temporal f i l ter
are adequately described by a second-order low-pass
filter with a time constant r ry 40 msec. Effectively, this
fi l ter corresponds to a temporal integration of the vel-
ocity signal of at least 100 msec. This is consistent with
the upper temporal l imit of about 100 msec for the
integration of velocity information (improving signal-to-noise ratios) in motion discrimination experiments(DeBruyn & Orban, 1988; Snowden & Braddick, 1991).
Thus, the lower and upper l imits for temporal inte-gration in the human visual motion system are equal,
suggesting a single hard-wired temporal filter in the
motion processing system. This view is further supportedby the close quantitative correspondence between the
increase of cut-offfrequency with speed (as reported here
for motion modulation detection) and the decrease of
temporal integration time with speed found in motion
discrimination studies (van Doorn & Koenderink, 1982,
l e85).This leads to the intriguing question: which stage in
the stream of visual motion processing accounts for the
characteristic temporal filtering found in our exper-
iments?
Temporal filtering: mechanistic considerations
A functional description of the phenomenology of our
experiments consists of a temporal integration of anunsmoothed internal representation of the true velocity
signal [Fig. 8(a)]. At this point we will try to link thisfunctional description to an actual implementation in thevisual system in terms of well-known motion detectionmechanisms.
An abstract description in terms of a smoothedmotion signal does not necessarily mean that the visual
PETER WERKHOVEN
f o u n d r ( @ r , 2 ) : 3 4 m s e c a n d
et al.
FIGURE 8. Two similar (but fundamentally different) modulation
detection processing streams. The first stream (a) consists of the
extraction of a true uelocitt' signal. followed by iow-pass temporal
fi ltering and by a final detection stage (e.g. peak or variance detection).
The second stream (b) consists of a velocity extraction that does not
yield a true velocity signal but a lov'-pass transformation of the true
t'elocitl ' . followed by a final detection stage as in the first stream.
system actually extracts (unfi ltered) velocity signal to
subsequentially low-pass fi l ter it in t ime. In flact, thefollowing rhetorical questions make such an implemen-
tation unlikely: (l) How does the visual system arrive at
the representation of the true (unfi ltered) velocity in the
first place? (2) lf such a representation exists. should this
signal be low-pass fi l tered given the great advantages of
having access to a velocity signal with high temporalresolution (Arnspang, 1988; Gli inder. 1990)? Because of
the above puzzles. we believe that the temporal inte-gration is inherent to the mechanism that arrives at a
velocity representation and that it takes place effectivelybefore the final estimate of the velocity vector. akin tothe scheme of Fig. 8(b).
We il lustrate some of the possible stages of temporalfi l tering by adopting a specific but plausible basic motion
detector: the Reichardt-correlator (van Doorn & Koen-
der ink. 1985), see Fig. 9.A plausible implementation of such a correlator typi-
cally contains three temporal f i l tering stages:
l. A temporal prefi lter f(t) for each input l ine.2. A temporal delay fi l ter (with time constant t ) in
one of the input l ines.3. A temporal low-pass fi l ter J. of the correlatorinput.We will discuss each of these fi l ters as candidates to
account for the temporal low-pass fi l tering of the vel-
ocity vector modulation functions found in our exper-iments.
Temporal lsvt-pass filtering of correlator outpur. Intui-
t ively, it is tempting to associate the psychophysically
observed intergration of velocity with the temporalintggration J. of the correlator output. Perhaps suripris-ingly, however, temporal f i l ter [, is not equivalent with
FIGURE 9. A standard motion detector (Reichardt correlator). Stan-
dard motion analysis consists of two input l ines (receptive fields) with
temporal f i l ters / ( l ) , a delay f i l ter (wi th t ime constant r ) for one of the
input l ines. a correlat ion stage and a temporal integrat ion f i l ter Jr
atemporal in tegrat ionof themodulat ionofspeedorupperr ightcornerofF ig. l0 .Tomakeourargumentasdirection. To show this, we will consider an ensemble of strong ai possible, we ?:::t:d :T::T,
temporal inte-
motion detectors (Reichardt "orr"iuto..),
ideally tuned gration t;kes place within a temporal window that
to a continuum of velocities. r. ,"lr"r1ii"i" -i, poini J*""ra, the period of the modulation function a few
we will focus on speed modulati"; f*;ii;;t Assuming times such that it flattens the profile'
that detecrors tuned to identical'J""iii*" "."
p."r"i Figure t0 also shows the ensemble activation profiles
(Glilnder, 1990), this ensemble "un
i" pu'"-at'i"a- ty for a unknown (unmodulated) speed function be/ore
tuning velocity u, only. At t,." r,-ittt 'ioning target hai (bottom left) and after temporal integration (bottom
velocity u(r) and wilt thus *"""," ".rr
i"i"!,;";* a iigt t). ouuiourty a ionstant profile in time is invariant
it"rtg'""r".r,y ,, = u(r1' Therefore'-t# type of activated under temporal integration'
detectors (parameterized Uy ,,; witiin the ensemble will As a result of temloral integration Jr (blurring along
vary in time, yielding time depende'nt )nsemble artiuation the horizontal time axis), both the ensemble activation
profiles. Forexample' ensemble ;;;t;;tt*';;;;;; for a profiles for modulated and unmodulated velocitv func-
iriangular speed modulation iun.iion ur"'giu.n in the tions. are constant it time' However' the shape of the
left upper corner of Fig. 10. ptofilet io' the modulated speed function (upper right)
In th is f igurethetypeofdetector(parameter izedbyi tsandthat for theunmodulatedspeedfunct ion(bot tomtunins speed t',, along the *"# #;;;;;
"tirtit"J righQ differ stronglv even for infinite blurring'
by the moving dot is given as u iun"tion oiti-" r iutong A true integrat'ion of the speed modulation signal'
the horizontal axis). wten tne oeiecio" utt u"'y 'hutply howeuer' tnou-id blut the ensemble profile along the
tuned, only one type or a.t."tor'i, -u.tiu"
oi u ti-", vertical6Deed) axis yielding blurred profiles that become
dependent on the sPeed "r,n.
.J"itg a"i ii";;;t;. indiscriminable for infinite blurring'
l a r m o m e n t i n t i m e , w e c a n w a l k a l o n g t h e v e r t i c a l a x i s l h e a b o v e r e a s o n i n g s h o w s t h a t t e m p o r a l l y f i l t e r i n gand find which detectors are active as a-function of their the speed sign-al is not iquivalent to-temporally filtering
tuning speed (the ensemble *U"",i"" pr"nf.l. Because the oulput "of
motion. d;teclors (correlaton) That is'
the dot moves with a single ;il;; !;;ffi;l"niti tr"mng in the speed dimension is g€n€rallv not equival-
rime. the ensemble profile is a single pulse thar shifts ent to ;lurring in _the time diT:1sio1-.we suggest that
along the verrical axis in tine.'- fritri sharply tuned the psychophisically observed integration of the speed
detectors. the ensemble activation profile is a perfect signal musi be inherent to a processing stage which
copy of the physical speed .oaotution signal u(i)' and comes before the correlation stage'
is thus triangular in time t"t'tpp"t ftft "o'n"'
of ,Tempiral
pre-fher' The shown temporal low-pass
F i g . l 0 ) ' c h a r a c i e r i s t i c f o i m o d u l a t i o n d e t e c t i o n m i g h t b e i n -N o w l e t . s c o n s i d e r t h e t e m p o l a l f i l t e r J r t h a t i n t e g - n " . " n t t o t h e t e m p o r a l f i l t e r / ( r ) a l . t h e i n p u t o f a
trares rhe output of rhe standard motio; detectors in standard motion d;tector (see Fig g)' The reasoning
time yielding an integration or-,r,'" in-r..Lr. u"tirution would be in terms of "window of visibility" arguments
p r o f i l e a l o n g t h e h o r i z o n t a l t i m e a x i s ' T h e r e s u l t i n g a s u s e d t o e x p l a i n t h e p e r c e i v e d e q u a l i t y o f a p p a r e n thorizontally blurred acrivatron^;;;i.; i io*n in th! motion with rial motion at adequate sampling frequen-
Prwrx! 'J
"i"t (wat'o"
"r o/ ' 1986: Burr et al'' 1986)' However' in
Bero,e Frterino Aner Firrerino fS:illm*:nf,i".-r."i: flt""::1ffiffi:l':.1l:ooral cut-off frequencies of 4-8 Hz' At these unrealistic
:'STi y"Hl iemporal cut-off frequencies' the motion system has not
I c"",l ,"."tt"d ootimal sensitivity (Burr & Ross' 1982)!
OPTIC ACCELERATION232'l
t p e e o s p e e o t e l l l p u l a l L t r L - ( - , r r r r v 9 s v r r v r v
t + .u.n ,.u.hed optimal sensitivity (Burr & Ross. 1982)!
T , ^ . , ^ . _ d o , ' , . q u . n ' t y , t h e t e m p o r a l p r e - f i l t e r e x p l a n a t i o n i sa - \ - / - - \ z { > l I i m p l a u i i b l e ., - + ^"!i{l:;!,!J!"{,{";i'y;JH:il11il'if 'TilHlirunins tunins to account for low-pass transformations of modulat ion
speed speeJ functions. The standard motion detector as sketched in
{ + Fig. 9 is optimally activated if an object traverses the
r - l+ sPatial inteI t t . - . - ^ -
' r ' Lo . noa . l ^ f t hc
: . , , > | - i l ; ; ; i l s i n t h e f i n i t e d e l a y t i m e t . T h e s p e e d o f t h ei I - - - r - - -
- , ^ : ^ ^ + ^ - . , ^ . l ^ n - , r c f h e
;bj;;, ;;y vary during its trajectory' as long as the. ' ^ : - ^ r - , + L . i ^ - ^ - . , l t o i n qrime
_ above conit.aini is satisfied. Intuitively, this results in a
FIGURE 10. Ensemble activatlon profi les as a funclion ol t lme lhe
ensemble of motion detectors considered rn this figure is Paramet€riz" :",Tp,":11-^11::aging (or temporal integration) of the
;;;;';;; the tr'rning velocitv r', lvertical axisl Activation speed.. tunctron'
profiles are given as a fun"tion tim" r (t'otiz'onJ axis) fo;a ffiangular Gliinder (1989) has recently presented an interesting
vetocirv modula[ion functior (upper lef! corner) and a constant speed mathematical analysis on this issue. His study focused on
A"-",i.ri".**.i. e;-irt igitirta" "" rrt"* rhe resulting activation ina qu.*ion of
-how velocity estimates through an
prohles for both functlons after temporal inteeration (lr) ot eacn
motion detector output in the ensemble (ie iitegtatioi along dlg €nS€mole of standard motion detectots depend on the
horizontal axrs) spatrar object function and on the impulse-response
function of the delay filter of the detector. For anensemble of bilocal correlators tuned to a continuum ofvelocity vectors. he showed that the estimated velocityfunction is the result of the convolution of the true(physical) velocity vector function with a time-invariantkernel which only depends on the integral function of theimpulse-response function. Hence, the estimated vel-ocity vector function is independent of the spatial objectfunction.
Ghinder's proof strongly supports our view that thephenomenological description of our results. in terms oflow-pass temporal f i l tering of the velocity vector func-tion, corresponds with a plausible implementation interms inherently non-ideal (realizable) band-pass delayfi lters in correlator detectors. Following this hypothesis,the cut-off frequencies for modulation detection dependinversely on delay value r: for longer delay values, thewidth of convolution kernel increases and yields strongertemporal blurring. Our finding that cut-off frequencieso)c are s l ight ly dependent on speed r .o[ra.(r .o)x1, f l35] thuslead to the conclusion that delay values r are speeddependent. Cut-off frequencies coc are expected to beinversely dependent on correlator delay r. This con-clusion corresponds closely to the empirical power func-t ion reported by van Doorn and Koender ink (19g2):z ocr ,o oao.
Speed dependence of cut-off frequenciesThe dependence of cut-off frequency on speed we
observe in Fig. 3 is consistent with what has beenreported in both psychophysical and electrophysiologi-cal l i terature: higher velocities correspond to somewhatfaster detectors.
However, this finding is at odds with the speeddependence of temporal velocity resolution obtained bySnowden and Braddick (1991). They found that rhecut-off frequency for speed modulation detecti on de -creases with increasing velocity in their experimentalset-up. This issue remains to be resolved by furtherexperimentation. The most noticeable difference betweenour experiments and those of Snowden and Braddickconcerns the spatial nature of the stimuli used. Snowdenand Braddick used a spatially extended random-dotpattern centered at the fovea. whereas we used a local-ized target moving at a trajectory with an approximatelyconstant eccentricity in our case.
Relation to sampled (apparent) motion experimentsApparent motion differs from continuous (smooth)
motion because it is characterized by a speed functionthat is modulated in time. Apparent motion is thus aspecial case of the modulated functions examined in thispaper. We wi l l d iscuss two studies (Watson et a l . , l9g6;Burr et al., 1986) that reported on the minimum tem_poral sampling frequency yielding perceptual equival_ence of apparent and real motion and compare themwith our study on the upper cut-off frequencies forvelocity modulation detection.
In Watson et al.'s stroboscopic paradigm the time_dependent speed is an i lr-defined iignar, butleriodic with
PETER WERKHOVEN er a/.
a frequency equal to the strobe frequency. They findminimal sampling frequencies (yielding perceptualequivalence of stroboscopic and real motion) that aremuch higher ( > 30 Hz) than the cur-off frequencies forvelocity modulation detection obtained in this study.However, this is probably due to the strong luminancecue at low velocities.
In Burr er al. 's "sample and hold" paradigm themoving dot is visible all the time and displaced stepwisein time. The time-dependent speed function )1r) is nowwell-defined (see Fig. l) and can be compared with theblock modulation functions used in our experimentswith 100% modulat ion ampl i tude. Figure I shows that"sample and hold" speed modulation function f(l)differs only in duty cycle from the block modulationfunction fl(r) at 100% modulation amplitude as used inour experiments. Furthermore. functions )Jr ) differ onlyin peak width f rom the tr iangular funct ions A(t) atl00o/o amplitude. However. the cut-off frequencies foundin th is paper (approx. 4Hz for t r iangular modulat ionfunctions &t lo : 2 degrsec. see Fig. 3) differ by at leasta factor of 4 from minimum sampling frequencies( l5-40 Hz, dependent on the drift rare) found by Burr elal. for this particular speed. We offer a number ofexplanations for this apparent discrepancy:
l . For a given average speed 1:n. the ampl i tude ofthe fundamental frequency in the ..sample
andhold" morion modulat ion funct ion )11) is muchlarger than for the triangular speed modulationfunct ion A(t)u,e used to obtain Fig. 3 ( the rat ioequals n214 = 2.5). This al lows for higher cut_offfrequencies in Burr's paradigm.
2. Burr et a l . do not ment ion v isual f ixat ion. Thestrong dependence of detection performance onpursuit eye movements was discussed in theMethod section. Our observers reported to haveno dfficulties in detecting modulations whentracking the mo', ' ing dot for conditions wheredetection .failed under visual f ixation.
3. We mentioned before that for modulation fre_quencies that far exceed the temporal l imit(2 Hz) for the detection of modulations by themotion system. thresholds are determined bycues orl/side the motion system. such as spatialcues. Thus it may be that experiments on theequivalence of apparent and real motion do notexclusively reveal the structure of the visualmot ion system.
Relation to experiments with controllecl eye molementsResults have been reported on frequency limits for
velocity modulation detection when a mouing referenceis provided (Funakawa, 1989) (contrary to our station-ar.v fixation dot). Providing a moving reference leads tocut-off frequencies (x25Hz) that are considerablyhigher than those obtained in our study. Interesting asthese results are, we believe them to be indicative of thetemporal resolution of visual subsystems concerned withthe spatial analysis of moving patterns. and not with the
2328
OPTIC .\CCELERATION
determination of velocity as such. We believe both types
of experiments (and visual subsystems) should be clearly
distinguished (although, of course, they may be inti-
mately intertwined).
CONCLUSION
In conclusion. human detection of velocity vector
modulations is not based on optic acceleration [thetemporal derivative of the velocity modulation functionu(r)1. The data presented in this paper strongly supportthe view that modulation detection consists of a ttariancedetection process, based on the magnitude of a low-passfilter transformation of the true modulation functionu(r). Effectively, the motion system integrates the vel-
ocity vector modulation signal for about 100 msec overt ime.
These results put severe constraints on viable theoriesaiming to explain human capacities in the extraction ofthree-dimensional environmental information frommotion paral lax cues (Nakayama, 1985a).
REFERENCES
Arnspang, J. (1988). Opt ic acceierat ion. Proceedings ICCV,364-373.
Burgess. A. & Ghandehar ian. H. (1984). Visual srgnal detect ion. I .
Ability to use phase information. Journal of the Opttcal Societv of
America. / 1. 900 905.
Burr . D. C. & Ross. J. (1982). Contrast sensi t iv i t - " - at h igh veloci t ies.
Vision Researt'h. 2:. 119484.Bur r . D . C . . Ross . J . & Mor rone . M. C . (1986) . Smooth and samp led
motion. Vision Research. 26. 643452.
DeBru -v -n . B . & Orban . G . A . (1988) . Human ve loc in 'and d i rec t i on
discr iminat ion measured wi th random dot pat terns. Vis ion Re-
search. 28. 1323-1335.van Doorn. A. J. & Koender ink. J. J . (1982). Temporal propert ies of
the visual detectabilit;- of moving spatial white noise . Erperintental
Brain Research. 15. 179-188.Fahle. M. & Westheimer. G. (1988). Local and global factors in
disparity detection of rows of points. Vision Research. :8. 171-178.
Funakawa. M. (1989). Spat io- temporal ly averaged posi t ions of mov-
ing objects. Spat ia l V ' is ior t . 1.275-285.
Glunder, H. (1990). Correlat ive veloci ty est imat ion: Visual mot ion
analysis. independent of object form. in arrays of velocity tuned
bilocal detectors. Journul of the Optical Societt' o.l ' America, A7,
l5-s -26_1.
Green. D. M. & Swets. J. A. (1966). Srgt?ul detet ' t ion theort 'and
pst 'chophvsics. New York: Wi ley.
Hick. W. E. (1950). The threshold for sudden changes in the veloci ty
of a seen object. Quarterl.t' Journal o./' Experintental Psvt'hologr, 2.
3 3 { 1 .Julesz. B. (1971). Foundat ior ts oJ ' t 'vc lopean percept ior t . Chicago. I l l .
Univers i tv of Chicago Press.
Koender ink, J. J . (1986). Opt ic f low. Vis ion Research, :6 ' 16l-180.
Koendennk, J. J. & van Doorn. A. J. (1978). Detectability of power
fluctuations of temporal visuai noise. Zrsion Research. 1& l9l-195.
Koender ink, J. J . , van Doorn A. J. & van de Gr ind, W. A. (1985).
Spatial and temporal parameters of motion detection in the periph-
eral visual field. Journal of the Optical Societ.r- of America. A2,
252-259.McKee, S. P. (1981). A local mechanism for differential velocity
detection. Vision Research, 21, 491-500.
McKee, S. P. (1983). The spatial requirements for fine stereoacuity.
Vision Research. 2-1. l9l-198.
Nakayama, K. ( 1985a). Higher order derivatives of the optical velocity
vector field: Limitations imposed by bioiogical hardware. In Igle.
D., Jeannerod, M. & Lee. D. (Eds), Brain mechanisms and spatial
uision. Holland: Martinus Ntjhoff.
Nakayama, K. (1985b). Bio logical image mot ion processing: A review.
Vision Research. 25, 625460.
Nakayama, K. & Tyler . C. W. (1981). Psychophysicai isolat ion of
movement sensitivity by removal of familiar position cues. Vision
Research. 21. 42'7_,{33.
Rashbass. C. (1970). The visibil ity of transient changes of lumtnance.
Journal of Physiolog.v. 210, 165-186.
Rashbass. C. (1976). Uni f icat ion of two contrast ing models of the
visual incremental threshold. l/ ision Reseurch, 16. l28l 1283.
Regan, D. M.. Kaufman, L. & Lincoln. J. (1986). Mot ion in depth and
visual accelerat ion. In Bof f . K. R. . Kaufman. L. & Thomas. J. P.
(Eds), Handbook of human perception and human perJormance
(Vol . I . Chap. l9) . New York: Wi ley.
Schwartz, B. J. & Kaufman. L. (1987). Psychophysical measures for
changes in direction of motion. Sot'ietv of Neuroscience Abstracts.
r 3 . 6 3 t .Snowden. R. J. & Braddick. O J (1991). The temporal integrat ion and
resolution of velocity signals. Visiort Research. 3l,. 907-914.
Watson. A.B (1987). Est imat ion of local spat ia i scaie. Journal oJ the
Optical Society o.f ' .4merica. '44. 1519-1582.
Watson , A . B . , Ahumada . A . J . & Far re l l . J . E . (1986) . W indow o f
v is ib i l i ty : A psychophvsical theory of f idel i ty in t ime-sampled v isual
motion displays. Journal of the Optical Societv o.f ' Anteric'a. .43.
300-307.Wat t . R . J . (1985) . Image segmenta t ion a t con tou r i n te rsec t ions in
human focal vision. Journal ol' the Optical Soc'ietv ol Anterica. A2.
I 200- I 204.Werkhoven. P. & Koender ink. J. J . (1991). Visual processing of rotary
motion. Perception and Psvchoph.r'sics. 49. 73-82.
Wertheim. T. (1894). Uber die indirekte Sehschdrfe. Zeitschri/ 't / i ir
Pst'chologie und Ph.t'siologie der Sinnesorgane. 7. 172 189.
Westheimer. G (1982). The spat ia l -erain of the per i foveal v isual f ie ld.
l'ision Reseurch. ::. 157 162"
Acknov'ledgements-This research was performed at the Institute for
Percept ion TNO. Soesterberg. The Nether lands. The research of Peter
Werkhoven was supported by the USAF Life Science Directorate.
Visual Informat ion Processing Grant 88-0140. Herman Snippe was
supported by the InSight project of the ESPRIT Basic Research
Actions of the European Community'. Alex Toet was supported by
NATO erant CRG 890970.