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
Stereo by Optimization Early algorithms: dynamic programming –(Baker ‘81, Belumeur & Mumford ‘92…) –Don’t generalize beyond 2 camera, single scanline
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
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)
Planar Graph Shortest Path Given: an embedded planar graph –faces, edges, vertices
Planar Graph Shortest Path A non negative cost on each edge 57
Planar Graph Shortest Path Two boundary points on the exterior of the complex
Planar Graph Shortest Path Find minimal curve: (collection of edges) with given boundary
Planar Graph For stereo Camera LeftCamera Right Selected Match Selected Occlusion
Algorithms Classic: Dijkstra’s –Works even for non-planar graphs Wacky: use duality –But this will generalize to higher dimension
Duality
face vertex edge cross edge - same cost 57
Duality Split exterior
Source Sink Source Duality Add source and sink
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
Source Sink Cuts Claim: Primalization of MIN-CUT will be shortest path
Sink Source Sink Why this works Cuts of dual graph = partitions of dual verts
Sink Source Sink Why this works Partition of dual verts = partition of primal faces
Sink Source Sink Source Sink Why this works Partition of primal faces = primal path
Sink Source Sink Source Sink Why this works Cuts in dual correspond to paths in primal MIN-CUT in dual corresponds to shortest path in primal
Same idea works for surfaces!
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
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
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
Dual construction for min surf face vertex edge cross edge Sink Source cell vertex face cross edge MIN-CUT primalizes to min surf
Checkpoint Solve for minimal paths and surfaces –MIN-CUT on dual graph Apply these algorithms to stereo vision
Flatland Stereo Camera LeftCamera Right Geometric interpretation of Cox et al. 96 pixel
Flatland Stereo Camera LeftCamera Right Geometric interpretation of Cox et al. 96 pixel
Flatland Stereo Camera LeftCamera Right Cost: unmatched/discontinuity, β
Flatland Stereo Camera LeftCamera Right Cost: correspondence quality
Flatland Stereo Camera LeftCamera Right
Camera LeftCamera Right Match Unmatched Uniqueness & monotonicity solution is directed path Flatland Stereo
Camera LeftCamera Right Match Occlusion, discontinuity Note: unmatched pixels also function as discontinuities Flatland Stereo
Flatland to Fatland Camera LeftCamera Right
Flatland to Fatland Camera LeftCamera Right
2 cameras, 3d
One Cuboid Among Many Solve for minimal surface
Geometric interpretation IG98
Three Camera Rectification (Ayache ‘88)
Three Camera
One cuboid
Dual graph of one cuboid
One Cuboid Among Many Solve for minimal surface
More divisions of middle cell
More expressive decomposition
Complexity Vertices and edges: 20 n d –n: pixels per image –d: max disparity Time complexity O((nd) 2 log(nd)) About 1 min
Results LL image RC KZ01 MS
LL image RC KZ01 MS
Future Application of MS to n cameras –Monotonicity/oriented manifold enforces more than uniqueness –see Kolmogorov & Zabih (today 11:00am) Other applications of MS