minimal surfaces for stereo chris buehler, steven j. gortler, michael f. cohen, leonard mcmillan...

54
Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Upload: maude-hensley

Post on 17-Dec-2015

217 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Minimal Surfaces for Stereo

Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan

MIT, HarvardMicrosoft Research, MIT

Page 2: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Motivation

• Optimization based stereo over greed based– No early commitment– Enforce interactions: each pixel sees unique item– Penalize interactions: non-smoothness

Page 3: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Stereo by Optimization

• Early algorithms: dynamic programming– (Baker ‘81, Belumeur & Mumford ‘92…)– Don’t generalize beyond 2 camera, single scanline

Page 4: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Stereo by Optimization

• Recent Algorithms: iterative expansion– (… Kolmogorov & Zabih ‘01)– very general– NP-Complete

• Local opt found quickly in practice

• Recent algorithms: MIN-CUT– (Roy & Cox ‘96, Ishikawa & Geiger ‘98) – Polynomial time global optimum– New interpretation to such methods

Page 5: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Contributions

• Stereo as a discrete minimal surface problem• Algorithms: Polynomial time globally optimal

surface– Using MIN-CUT (Sullivan ‘90)– Build from shortest path

• Applications to stereo vision– Rederive previous MIN-CUT stereo approaches– New 3-camera stereo formulation (Ayache ‘88)

Page 6: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Planar Graph Shortest Path

• Given: an embedded planar graph– faces,

edges, vertices

Page 7: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Planar Graph Shortest Path

• A non negative cost on each edge

57

Page 8: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Planar Graph Shortest Path

• Two boundary points on the exterior of the complex

Page 9: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Planar Graph Shortest Path

• Find minimal curve: (collection of edges) with given boundary

Page 10: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Planar Graph For stereo

Camera Left Camera Right

Selected MatchSelected Occlusion

Page 11: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Algorithms

• Classic: Dijkstra’s– Works even for non-planar graphs

• Wacky: use duality– But this will generalize to higher dimension

Page 12: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Duality

Page 13: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Duality

• face vertex • edge cross edge

- same cost57

Page 14: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Duality

• Split exterior

Page 15: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Source

Sink

Source

Duality

• Add source and sink

Page 16: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Cuts

Source

Sink

• Cuts of dual graph = partitions of dual verts• Cost = sum of dual edges spanning the partition• MIN-CUT can be found in polynomial time

Page 17: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Source

SinkCuts

• Claim: Primalization of MIN-CUT will be shortest path

Page 18: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Sink

SourceSource

SinkWhy this works

• Cuts of dual graph = partitions of dual verts

Page 19: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Sink

SourceSource

SinkWhy this works

• Partition of dual verts = partition of primal faces

Page 20: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Sink

SourceSource

Sink

Source

SinkWhy this works

• Partition of primal faces = primal path

Page 21: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Sink

SourceSource

Sink

Source

SinkWhy this works

• Cuts in dual correspond to paths in primal

• MIN-CUT in dual corresponds to shortest path in primal

Page 22: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Same idea works for surfaces!

Page 23: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Increasing the dimension

Planar graph: verts, edges, faces cost on edges boundary: 2 points on exterior sol: min path

Spacial compex: verts, edges, faces, cells cost on faces boundary: loop on exterior sol: min surface

Page 24: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Increasing the dimension

Planar graph: verts, edges, faces boundary: 2 points on exterior sol: min path

Spacial compex: verts, edges, faces, cells cost on faces boundary: loop on exterior sol: min surface

Page 25: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Increasing the dimension

Planar graph: verts, edges, faces boundary: 2 points on exterior sol: min path

Spacial compex: verts, edges, faces, cells cost on faces boundary: loop on exterior sol: min surface

Page 26: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Dual construction for min surf

• face vertex • edge cross edge

Sink

Source

• cell vertex • face cross edge

MIN-CUT primalizes to min surf

Page 27: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Checkpoint

• Solve for minimal paths and surfaces– MIN-CUT on dual graph

• Apply these algorithms to stereo vision

Page 28: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Flatland Stereo

Camera Left Camera Right

Geometric interpretation of Cox et al. 96

pixel

Page 29: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Flatland Stereo

Camera Left Camera Right

Geometric interpretation of Cox et al. 96

pixel

Page 30: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Flatland Stereo

Camera Left Camera Right

Cost: unmatched/discontinuity, β

Page 31: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Flatland Stereo

Camera Left Camera Right

Cost: correspondence quality

Page 32: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Flatland Stereo

Camera Left Camera Right

Page 33: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Camera Left Camera Right

MatchUnmatched

Uniqueness & monotonicity solution is directed pathFlatland Stereo

Page 34: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Camera Left Camera Right

MatchOcclusion, discontinuity

Note: unmatched pixels also function as discontinuities

Flatland Stereo

Page 35: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Flatland to Fatland

Camera Left Camera Right

Page 36: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Flatland to Fatland

Camera Left Camera Right

Page 37: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

2 cameras, 3d

Page 38: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

2 cameras, 3d

Page 39: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

One Cuboid Among Many

Solve for minimal surface

Page 40: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Geometric interpretation IG98

Page 41: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Three Camera

Rectification (Ayache ‘88)

Page 42: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Three Camera

Page 43: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Three Camera

Page 44: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Three Camera

Page 45: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Three Camera

Page 46: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

One cuboid

Page 47: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Dual graph of one cuboid

Page 48: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

One Cuboid Among Many

Solve for minimal surface

Page 49: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

More divisions of middle cell

Page 50: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

More expressive decomposition

Page 51: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Complexity

• Vertices and edges: 20 n d– n: pixels per image– d: max disparity

• Time complexity O((nd)2 log(nd))

• About 1 min

Page 52: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

ResultsLL image

RC

KZ01

MS

Page 53: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

LL image

RC

KZ01

MS

Page 54: Minimal Surfaces for Stereo Chris Buehler, Steven J. Gortler, Michael F. Cohen, Leonard McMillan MIT, Harvard Microsoft Research, MIT

Future

• Application of MS to n cameras– Monotonicity/oriented manifold enforces more

than uniqueness– see Kolmogorov & Zabih (today 11:00am)

• Other applications of MS