Grab Cut in One Cut Meng Tang Lena
Grab. Cut in One Cut Meng Tang Lena Gorelick Olga Veksler Yuri Boykov http: //vison. csd. uwo. ca Basic Segmentation Energy L 1 Color Separation Term • Graph Cut [1] fixes appearances q q 0, 1 • Grab. Cut [2] optimizes appearances q q 0, • L 1 penalty between color histograms of foreground and background: ai entropy of colors in low entropy example S S low entropy => color separation • Decomposing energy [3] nonsubmodular Bounding box Seeds One Cut results • Overview of graph structure Template No L 1 With L 1 Hamming distance between template and segment + L 1 color separation Jensen-Shannon (JS) color separation submodular [4] L 1 color separation Replace JS with L 1 (ballooning, saliency, seeds, etc. ) [1] Y. Boykov and M. Jolly, Interactive Graph Cuts for Optimal Boundary & Region Segmentation of Objects in N-D Images, ICCV 2001. One Cut result Application – Shape Matching and Saliency Segmentation Input Main New Ideas To avoid trivial solution, replace with application dependent unary term hard constrains + L 1 color separation simpler than construction in [4] S S With strokes/seeds: heuristic foreground ballooning + L 1 color separation • Our construction: add one auxiliary node ai for each color bin i high entropy example Grab Cut volume balancing With bounding boxes: 1 • Our interpretation via entropy: entropy of colors in S Application – Interactive Segmentation TP FP Time No L 1 76. 97% 6. 96% 4. 1 s With L 1 81. 88% 3. 86% 13. 0 s connect pixels of each color to corresponding auxiliary nodes Contributions • L 1 penalty for color appearance overlap between segments n • Simpler construction wrt P -Potts [4] in a new context (color separation rather then superpixel consistency) • Proposing various practical unary terms instead of GC volume balancing Input Saliency No L 1 With L 1 Saliency-based unary terms + L 1 color separation GT [2] C. Rother, V. Kolmogorov and A. Blake, Grab. Cut: Interactive Foreground [3] S. Vicente, V. Kolmogorov and C. Rother, Joint Optimization [4] P. Kohli, L. Ladicky, and P. H. S. Torr. Robust Higher Order Potentials Extraction using Iterated Graph Cuts, SIGGRAPH 2004. of Segmentation and Appearance Models, ICCV 2009. for Enforcing Label Consistency. IJCV, 2009.
- Slides: 1