1. Groups of Adjacent Contour Segments for Object Detection Vittorio Ferrari Loic Fevrier Frederic Jurie Cordelia Schmid

2. Problem: object class detection & localization Training Testing ? Focus: classes with characteristic shape

3. Features: pairs of adjacent segments (PAS) Contour segment network [Ferrari et al. ECCV 2006] 1)edgels extracted with Berkeley boundary detector 2) edgel-chains partitioned into straight contour segments 3) segments connected at edgel- chains’ endpoints and junctions

4. Features: pairs of adjacent segments (PAS) segments connected in the network PAS = groups of two connected segments encodes geometric properties of the PAS scale and translation invariant compact, 5D PAS descriptor:

5. Features: pairs of adjacent segments (PAS) Example PAS Why PAS ? + intermediate complexity: good repeatability- informativeness trade-off + scale-translation invariant + connected: natural grouping criterion (need not choose a grouping neighborhood or scale) + can cover pure portions of the object boundary

6. PAS codebook Based on descriptors, cluster PAS into types a few of the most frequent types based on 10 outdoor images (5 horses and 5 background). types based on 15 indoor images (bottles) Frequently occurring PAS have intuitive, natural shapes As we add images, number of PAS types converges to just ~100 Very similar codebooks come out, regardless of source images + general, simple features. We use a single, universal codebook (1 st row) for all classes

7. Window descriptor 1. Subdivide window into tiles. 2. Compute a separate bag of PAS per tile 3. Concatenate these semi-local bags [Lazebnik et al. CVPR 2006]; [Dalal and Triggs CVPR 2005] + distinctive: records which PAS appear where weight PAS by average edge strength + flexible: soft-assign PAS to types rather coarse tiling + fast to compute using Integral Histograms

8. Training 1. Learn mean positive window dimensions 2. Determine number of tiles T 3. Collect positive example descriptors 4. Collect negative example descriptors: slide window over negative training images

9. Training 5. Train a linear SVM Here a few of the top weighted descriptor vector dimensions (= 'PAS + tile'): + lie on object boundary (= local shape structure common to many training examples)

10. Testing 1. Slide window of aspect ratio, at multiple scales 2. SVM classify each window + non-maxima suppression detections

11. Results – INRIA horses + tiling brings a substantial improvement optimum at T=30 -> keep this setting on all other experiments + works well: 86% det-rate at 0.3 FPPI (with 50 pos + 50 neg training images) Dataset: ~ Jurie and Schmid, CVPR 2004 170 positive + 170 negative images (training = 50 pos + 50 neg) wide range of scales; clutter (missed and FP)

12. Results – INRIA horses Dataset: ~ Jurie and Schmid, CVPR 2004 170 positive + 170 negative images (training = 50 pos + 50 neg) wide range of scales; clutter + PAS better than any IP all interest point (IP) comparisons with T=10, and 120 feature types, (= optimum over INRIA horses, and ETHZ Shape Classes; all IP codebooks are class-specific) (missed and FP)

13. Results – Weizmann-Shotton horses Dataset: Shotton et al., ICCV 2005 327 positive + 327 negative images (training = 50 pos + 50 neg) no scale changes; modest clutter Shotton’s EER - exact comparison to Shotton et al.: use their images and search at a single scale - PAS same performance (~92% precision-recall EER), but: + no need for segmented training images (only bounding-boxes) + can detect objects at multiple scales (see other experiments)

14. Results – ETHZ Shape Classes Dataset: Ferrari et al., ECCV 2006 255 images, over 5 classes training = half of positive images for a class + same number from the other classes (1/4 from each) testing = all other images large scale changes; extensive clutter

15. Results – ETHZ Shape Classes Dataset: Ferrari et al., ECCV 2006 255 images, over 5 classes training = half of positive images for a class + same number from the other classes (1/4 from each) testing = all other images large scale changes; extensive clutter Missed

16. Results – ETHZ Shape Classes + mean det-rate at 0.4 FPPI = 79% + PAS >> I.P for apple logos, bottles, mugs PAS ~= IP for giraffes (texture!) PAS < IP for swan + overall best IP: Harris-Laplace + class specific IP codebooks GiraffesMugsSwans Apple logosBottles

17. Results – Caltech 101 Dataset: Fei-Fei et al., GMBV 2004 42 anchor, 62 chair, 67 cup images train = half + same number of caltech101 background testing = other half pos + same number of background scale changes; only little clutter

18. Results – Caltech 101 Dataset: Fei-Fei et al., GMBV 2004 On caltech101’s anchor, chair, cup: + PAS better than Harris-Laplace + mean PAS det-rate at 0.4 FPPI: 85%

19. Comparison to Dalal and Triggs CVPR 2005 GiraffesMugsSwans Apple logos Bottles

20. Comparison to Dalal and Triggs CVPR 2005 Caltech anchorsCaltech chairsCaltech cups INRIA horses Shotton horses + overall mean det-rate at 0.4 FPPI: PAS 82% >> HoG 58% PAS >> HoG for 6 datasets PAS ~= HoG for 2 datasets PAS < HoG for 2 datasets

21. Generalizing PAS to kAS kAS: any path of length k through the contour segment network segments connected in the network 3AS 4AS scale+translation invariant descriptor with dimensionality 4k-2 k = feature complexity; higher k -> more informative, but less repeatable kAS overall mean det-rates (%) 1AS PAS 3AS 4AS 0.3 FPPI 69 77 64 57 0.4 FPPI 76 82 70 64 PAS do best !

22. Conclusions Connected local shape features for object class detection Experiments on 10 diverse classes from 4 datasets show: + better suited than interest points for these shape-based classes - fixed aspect-ratio window: sometimes inaccurate bounding-boxes + object detector deals with clutter, scale changes, intra-class variability - single viewpoint + PAS have the best intermediate complexity among kAS + object detector compares favorably to HoG-based one

23. Current work: detecting object outlines Training: learn the common boundaries from examples Model collection of PAS and their spatial variability only common boundary

24. 1. detect edges Current work: detecting object outlines Detection on a new image 2. match PAS based on descriptors 3. vote for translation + scale initializations 4. match deformable thin-plate spline based on deterministic annealing Outline object in test image, without segmented training images !

25. A few preliminary results

26. Results – Caltech 101 Dataset: Fei-Fei et al., GMBV 2004 On caltech101’s anchor, chair, cup: + PAS better than any IP + mean PAS det-rate at 0.4 FPPI: 85%