1 / 27

Category Independent Region Proposals

Category Independent Region Proposals. Ian Endres and Derek Hoiem University of Illinois at Urbana-Champaign. Finding Objects. Scanning Window. Horse Dog Cat Car Train …. 10,000+ windows. Category Independent Search. ~100 regions. Finding Unfamiliar Objects. Finding Objects.

spike
Télécharger la présentation

Category Independent Region Proposals

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Category Independent Region Proposals Ian Endres and Derek Hoiem University of Illinois at Urbana-Champaign

  2. Finding Objects

  3. Scanning Window Horse Dog Cat Car Train … 10,000+ windows

  4. Category Independent Search ~100 regions

  5. Finding Unfamiliar Objects

  6. Finding Objects Objectives: • Minimize number of proposed regions • Maintain high recall of all objects • Provide detailed spatial support (i.e. segmentation)

  7. Challenges • Objects extremely diverse • Variety of shapes, sizes • Many different appearances • Within object variation • Multiple materials and textures • Strong interior boundaries • Many objects in an image

  8. Overview Generate Proposals: Maximize recall Rank Proposals: Small diverse set of object regions 1 2 3 4 ...

  9. Generating Proposals 1. Select Seed 2. Compute affinities for seed 5. Change parameters Repeat 3. Construct binary CRF 4. Compute proposal + Unary term: Affinities Pairwise term: Occlusion Boundaries

  10. Generating Seeds • Compute occlusion boundaries (Hoiem et al. ICCV ‘07) • Generate hierarchal segmentation • Incrementally merge regions of oversegmentation • Use regions with sufficient size and boundary strength • Avoids redundant or uninformative seeds

  11. Region Affinity • Learned from pairs of regions belonging to an object • Computed between the seed and each region of the hierarchy • Features: color and texture similarity, boundary crossings, layout agreement

  12. Color/Texture Similarity • Color, texture histograms for each region • Compute histogram intersection distance between two regions

  13. Boundary Crossing • Draw line between region centers of mass • Compute strength of occlusion boundaries crossed

  14. Layout Agreement • Predict object extent from each region • Compute strength of agreement between two regions

  15. CRF Segmentation • Binary segmentation • Graph composition: • Nodes: Superpixels • Edges: Adjacent superpixels +

  16. CRF Segmentation • Graph Potentials • Unary Potential: affinity values for each superpixel • Edge Potential: occlusion boundary strength • Parameters (25 combinations) • Node/Edge weight tradeoff • Node bias + Unary potential: Affinities Edge potential: Occlusion Boundaries

  17. Ranking Proposals Generated Ranking Appearance scores 1. wT X1 wT X2 Sort scores 2. wT X3 3. wT X4 4.

  18. Lacks Diversity • But in an image with many objects, one object may dominate 1 … 20 2 … 50 … 3 100 … 150 4

  19. Encouraging Diversity • Suppress regions with high overlap with previous proposals … 1 20 2 … 3 50 4 … … 100 10

  20. Ranking as Structured Prediction • Find the max scoring ordering of proposals • Greedily add proposals with best overall score Appearance score Overlap penalty Gives higher weight to higher ranked proposals Overall score

  21. Learning to Rank(Max-margin Structured Learning) • Score of ground truth ordering (R(n)) should be greater than all other orderings (R): • Loss ( ) encourages good orderings: • Higher quality proposals should have higher rank • Each object should have a highly ranked proposal

  22. Experimental Setup • Train on 200 BSDS images • Test 1: 100 BSDS images • Test 2: 512 Images from Pascal 2008 Seg. Val.

  23. Evaluation • Region overlap • Recall at 50% region overlap • Typically more strict that 50% bounding box overlap • Measures detection quality and segment quality Ai Aj

  24. Qualitative Results BSDS (Rank, % overlap) Pascal

  25. Vs. Standard Segmentation Ours: 80% 180 proposals Standard: 80% 70,000 proposals (merge 2 adjacent regions) Standard: 53% 3000 proposals Ours: 53% 18 proposals

  26. Recalling Pascal Categories

  27. Future work • Object Discovery • Incorporate into detection systems • Label regions directly • Voting from proposed regions • Refine proposals with domain knowledge • i.e. wheel or head models

More Related