coil departmen computer t rep ort no cucs i in tro duction w eha v e constructed a database of color...
TRANSCRIPT
Columbia Object Image Library �COIL�����
Sameer A� Nene and Shree K� Nayar and Hiroshi Murase
Department of Computer Science
Columbia University
New York� N�Y� �����
sameer�cs�columbia�edu
nayar�cs�columbia�edu
Technical Report No� CUCS������
Abstract
Columbia Object Image Library �COIL����� is a database of color images of ��� objects�The objects were placed on a motorized turntable against a black background� The turntablewas rotated through ��� degrees to vary object pose with respect to a �xed color camera�Images of the objects were taken at pose intervals of degrees� This corresponds to � posesper object� The images were size normalized� COIL���� is available online via ftp�
i
� Introduction
We have constructed a database of ���� color images of ��� objects �� images per object��The objects have a wide variety of complex geometric and re ectance characteristics �see�gure ��a��� The database� called Columbia Object Image Library �COIL������ was used ina real�time ��� object recognition system �Nayar et al������b� �Nayar et al������c�� Figure��b� shows an object �a toy cat� from the database being placed in front of the system sensor�Figure ��c�� shows the recognized object and it�s pose in the upper right corner as displayedby the system� The recognition system uses the parametric eigenspace technique �Muraseand Nayar����� for visual learning and recognition� For related publications� see �Nayarand Poggio������ �Nayar et al������a� �Nene and Nayar������� COIL���� is available by ftpfor research purposes �see Section ���
� Database Acquisition
The experimental setup used for image acquistion is shown in �gure �� A CCD color camerawith a �mm lens was �xed to a rigid stand about � feet from it�s base� A motorizedturntable was placed about � feet from the base of the stand� The camera was tilted downat about � degrees to point towards the turntable� This way most objects appeared at thecenter of the image when placed at the center of the turntable� To avoid strong shadows�only ambient � uorescent� room lighting was used� The camera was not very sensitive� so wehad to set the aperture to f��� �fully open�� A black background was provided by coveringthe turntable and visible background surfaces with black cloth�
Each object was placed in a stable con�guration at approximately the center of theturntable� The turntable was then rotated through ��� degrees and � images were takenper object� one at every degrees of rotation� Images were digitized using a DEC J��� colorframe grabber� Due to disk space limitations� it was impossible to store ���� color imagesof size ���x���� Moreover� for the ��� object recognition system mentioned above �Nayar etal������b� �Nayar et al������c�� it was necessary to normalize the size of an input image to���x���� Hence� for every image� we clip out the object from the black background using arectangular bounding box and resize it to ���x��� using interpolation�decimation �lters tominimize aliasing �Oppenheim and Schafer������� When resizing� aspect ratio is preserved�
In addition to size normalization� every image was histogram stretched� i�e� the intensityof the brightest pixel was made �� and intensities of the other pixels were scaled accord�ingly� The images were saved as ���bit PPM � �portable pixmap� color images� Note thatPPM images can be viewed with xv Ver� ���� A sample �lename of a database image is�obj ���ppm�� The pre�x obj identi�es the object� The numeric value �� following thedouble underscore separator indicates the pose in degrees� The su�x �ppm indicates the �letype� The database is available as a single compressed tar �le of size ��Mb� A grayscaleimage database COIL���� similar to COIL����� is also available� See �Nene et al��������
����bit PPM images contain data in little�endian format� When writing a program to read such PPM
images� care is needed to ensure that a possible con�ict with the machine endian is properly resolved�
�
�a�
�b� �c�
Figure �� The Columbia Object Image Library �COIL����� contains ���� images of ���objects� �a� The objects have a wide variety of complex geometric and re ectance charac�teristics� A real�time ��� object recognition system was constructed using COIL����� �b� Atoy cat from the database is shown to the system for recognition� �c� The system recognizesthe object in less than one second� The recognized object and it�s pose are displayed in theupper right hand corner�
�
Figure �� The objects were placed at the center of a motorized turntable� The turntable wasrotated through ��� degrees� An image was acquired with a �xed color camera at every degrees of rotation�
� Access Instructions
COIL���� is available by ftp over internet� All accesses are logged to help us know who isusing the database� The following is a sequence of commands required to download COIL�����
� ftp zen�cs�columbia�edu
Name� coil����
Password� Coil����
ftp� cd coil����
ftp� bin
ftp� get coil�����tar�gz
ftp� quit
In case of any problem or questions� the reader is advised to send mail to sameer�cs�columbia�eduor nayar�cs�columbia�edu�
Acknowledgements
This database was collected at the Center for Research on Intelligent Systems at the Depart�ment of Computer Science� Columbia University� It was supported by DOD�ONR MURIGrant N�������������� and a NSF National Young Investigator Award�
�
References
�Murase and Nayar� ���� H� Murase and S� K� Nayar� Visual Learning and Recognitionof �D Objects from Appearance� International Journal of Computer Vision� ����������January ����
�Nayar and Poggio� ����� S� K� Nayar and T� Poggio� Early Visual Learning� In S� K� Nayarand T� Poggio� editors� Early Visual Learning� Oxford University Press� March �����
�Nayar et al�� ����a� S� K� Nayar� H� Murase and S� A� Nene� Parametric Appearance Repre�sentation� In S� K� Nayar and T� Poggio� editors� Early Visual Learning� Oxford UniversityPress� March �����
�Nayar et al�� ����b� S� K� Nayar� S� A� Nene and H� Murase� Real�Time ��� Object Recog�nition System� In Proceedings of ARPA Image Understanding Workshop� Palm Springs�February �����
�Nayar et al�� ����c� S� K� Nayar� S� A� Nene and H� Murase� Real�Time ��� Object Recog�nition System� In Proceedings of IEEE International Conference on Robotics and Automa�tion� Minneapolis� April �����
�Nene and Nayar� ����� S� A� Nene and S� K� Nayar� SLAM� A Software Library for Ap�pearance Matching� In Proceedings of ARPA Image Understanding Workshop� Monterey�November ����� Also Technical Report CUCS��������
�Nene et al�� ����� S� A� Nene� S� K� Nayar and H� Murase� Columbia Object Image Library�COIL���� Technical Report CUCS������� Department of Computer Science� ColumbiaUniversity� February �����
�Oppenheim and Schafer� ����� A� V� Oppenheim and R� W� Schafer� Discrete�Time SignalProcessing� chapter �� pages �������� Prentice Hall� �����
�