160 likes | 184 Views
This study explores combining segmentation and recognition techniques to recover human body configurations from input images, addressing challenges like a variety of poses, clothing, missing parts, small support, and background clutter. The approach involves segmenting images into superpixels and assembling partial configurations based on torso and limb detectors. By using global constraints and pruning techniques, it refines and completes configurations to accurately localize joint positions in diverse poses. Results demonstrate the success of this approach in achieving accurate human figure segmentation despite imperfect initial segmentations.
E N D
Recovering Human Body Configurations: Combining Segmentation and Recognition Greg Mori, Xiaofeng Ren, Alyosha Efros and Jitendra Malik
Problem Input image Stick figure Support masks Single image: No background subtraction
Human Figures in Still Images • Detection of humans is possible for stereotypical poses • Standing • Walking (Mohan, Papageorgiou, and Poggio PAMI 01; Viola, Jones and Snow ICCV 03) • But we want to do more • Wider variety of poses • Localize joint positions
Why Is This Hard? • Variety of poses • Clothing • Missing parts • Small support for parts • Background clutter
Previous Work • Exemplar approaches • Shape Contexts (Mori and Malik ECCV 2002) • Order Structure (Sullivan and Carlson ECCV 2002) • LSH (Shakhnarovich et al. ICCV 2003) • Part-based approaches • Rectangles (Ioffe and Forsyth, IJCV 2001) • Corners (Song et al., PAMI 2003) • Dynamic programming (Felzenszwalb and Huttenlocher, CVPR 2000)
Approach:Unifying Segmentation and Recognition • Bottom-up • Detect half-limbs and torsos • Top-down • Assemble parts into human figure
SEGMENTS • Segmentation • How many? SUPERPIXELS Why Segmentation for Recognition? • Window-scanning (e.g. face detection)
Limb/Torso Detectors • Learn limb and torso detectors from hand-labeled data • Cues: • Contour • Average edge strength on boundary • Shape • Similarity to rectangle • Shading • x,y gradients, blurred • Focus • Ratio of high to low frequency energies
Islands of Saliency • “Partial configurations” • 3 half-limbs plus a torso • Combinatorial search over sets of limbs and torsos configurations
Pruning Partial Configurations • Many partial configurations are physically impossible • Prune using global constraints (not a tree) • Proximity • Relative widths • Maximum lengths • Symmetry in clothing colour • Results in 1000 partial configurations
Completing Configurations • Use superpixels to complete half-limbs • 2 or 3-limbed people • Sort partial configurations • Use limb, torso, and segmentation scores • Extend missing limb(s) of best configurations
LU ARM RL ARM LL LEG RL LEG LL LEG RL LEG
Results II Rank 3 Rank 3
Addressing the Challenges • Variety of poses • Parts-based approach, no restrictive priors • Clothing • Shape, shading, edge cues • Missing parts • Start with “islands of saliency” • Small support for parts • Start with “islands of saliency” • Background clutter • Global constraints on relative sizes of parts, colour
All segmentations are wrong, but some segmentations are useful!