A Globally Optimal Algorithm for Robust TV-L 1 Range Image Integration Christopher Zach VRVis Research Center Thomas Pock, Horst Bischof Institute for Computer Graphics and Vision TU Graz
Christopher Zach 2Robust Range Image Integration – 11/20/2016 Dense 3D Reconstruction E.g. site reconstruction from images Standard SfM
Christopher Zach 3Robust Range Image Integration – 11/20/2016 Dense 3D Reconstruction Multi-view stereo for dense depth Only statue model is desired No fg/bg segmentation “Clutter” in the depth maps
Christopher Zach 4Robust Range Image Integration – 11/20/2016 Depth Map/Range Image Fusion “Robust” GPU-based VRIP Proposed method: ☺ ?
Christopher Zach 5Robust Range Image Integration – 11/20/2016 Outline Introduction to TV and L 1 TV-L 1 Approach to Range Image Integration Results Conclusion/Outlook
Christopher Zach 6Robust Range Image Integration – 11/20/2016 Classification [1] Lempitsky & Boykov, “Global optimization for shape fitting” [2] Vogiatzis et al., “Multi-view stereo via volumetric graph cuts” [3] Hornung & Kobbelt, “Robust reconstruction of watertight models from non- uniformely sampled point clouds” [4] Faugeras & Keriven, “Variational principles, surface evolution, PDEs, level set methods and the stereo problem” [5] Kazhdan et al., “Poisson surface reconstruction”
Christopher Zach 7Robust Range Image Integration – 11/20/2016 TV + L 1 Fact Sheet Total Variation: L 1 data fidelity: TV-L 1 energy: E is convex, but not strictly convex (minimum not unique) For binary f: optimize for real-valued u, then apply thresholding e.g. iso-surface extraction in our domain More regularization: small features disappear instead of increased blurring [6] Chan & Esedoglu, “Aspects of TV Regularized L 1 Function Approximation”
Christopher Zach 8Robust Range Image Integration – 11/20/2016 Binary TV-L 1 Example
Christopher Zach 9Robust Range Image Integration – 11/20/2016 Multiple Inputs to Approximate Proposed TV-L 1 energy for multiple sources f i Simultaneously approximating all f i and minimize area of respective iso-surface while Inputs may have missing data Spatially regularized median λ may be spatially varying e.g. based on sampling density
Christopher Zach 10Robust Range Image Integration – 11/20/2016 Application to Depth Map Fusion Convert source depth maps to directional signed + truncated 3D distance fields Perform integration using the TV-L 1 approach Extract the final iso-surface from the fused result
Christopher Zach 11Robust Range Image Integration – 11/20/2016 How to Optimize E Direct optimization of E is difficult (or slow) Strictly convex energy functional u approximates v using the ROF energy v approximates u and the given data point-wise
Christopher Zach 12Robust Range Image Integration – 11/20/2016 Determine u for fixed v Problem: This is the Rudin-Osher-Fatemi energy [7] We employ the algorithm by Chambolle [8]: [7] Rudin et al., “Nonlinear Total Variation based Noise Removal Algorithms” [8] Chambolle, “An Algorithm for Total Variation Minimization and Applications” |p| ≤ 1
Christopher Zach 13Robust Range Image Integration – 11/20/2016 Determine v for fixed u Problem: This can be solved separately for every pixel/voxel Linear-time search through intervals If only missing values at x: See paper for details
Christopher Zach 14Robust Range Image Integration – 11/20/2016 Implementation Notes Signed distance generation on the GPU Encoding of input volumes “Run-length” compression of voxel data number of empty/occluded/missing entries sorted sequence of distance values e.g. 54 views, 360x400x240 voxel space: 380 MB instead of 6.9 GB Fusion accelerated by this scheme Coarse-to-fine for faster fill-in
Christopher Zach 15Robust Range Image Integration – 11/20/2016 Middlebury Multi-View Eval. Depth maps from GPU-based plane-sweep
Christopher Zach 16Robust Range Image Integration – 11/20/2016 Varying λ
Christopher Zach 17Robust Range Image Integration – 11/20/2016 Reducing # of Depth Maps 40 depth maps λ = depth maps λ = depth maps λ = depth maps λ = 1.2
Christopher Zach 18Robust Range Image Integration – 11/20/2016 Line-of-Sight Handling
Christopher Zach 19Robust Range Image Integration – 11/20/2016 More Results
Christopher Zach 20Robust Range Image Integration – 11/20/2016 Conclusion & Future Work Pros: Low quality depth maps yield high quality 3D meshes Easy to implement (numerical part is ~150 basic C++ LoC) Run-time in the magnitude of minutes Parallelized implementations on multi-core CPUs possible Yields global optimum Todos: Large scenery/high details e.g. sector-based fusion Incremental model update Photo-flux for direct multi-view stereo
Christopher Zach 21Robust Range Image Integration – 11/20/2016 Outlook Initial implementation on the GPU (Geforce 8800 GTX) 22 s 24 s 22 s
Christopher Zach 22Robust Range Image Integration – 11/20/2016 Thank you for your attention!