Princeton Shape Benchmark: Comparison of 3D Model Shape Descriptors

1. The Princeton Shape BenchmarkPhilip Shilane, Patrick Min, Michael Kazhdan, and Thomas Funkhouser

2. Shape Retrieval Problem 3D Model ShapeDescriptor BestMatches Model Database

3. Example Shape Descriptors • D2 Shape Distributions • Extended Gaussian Image • Shape Histograms • Spherical Extent Function • Spherical Harmonic Descriptor • Light Field Descriptor • etc.

4. Example Shape Descriptors • D2 Shape Distributions • Extended Gaussian Image • Shape Histograms • Spherical Extent Function • Spherical Harmonic Descriptor • Light Field Descriptor • etc. How do we know which is best?

5. Typical Retrieval Experiment • Create a database of 3D models • Group the models into classes • For each model: • Rank other models by similarity • Measure how many models in the same class appear near the top of the ranked list • Present average results

6. Typical Retrieval Experiment • Create a database of 3D models • Group the models into classes • For each model: • Rank other models by similarity • Measure how many models in the same class appear near the top of the ranked list • Present average results

7. Typical Retrieval Experiment • Create a database of 3D models • Group the models into classes • For each model: • Rank other models by similarity • Measure how many models in the same class appear near the top of the ranked list • Present average results

8. Typical Retrieval Experiment • Create a database of 3D models • Group the models into classes • For each model: • Rank other models by similarity • Measure how many models in the same class appear near the top of the ranked list • Present average results Query

9. Typical Retrieval Experiment • Create a database of 3D models • Group the models into classes • For each model: • Rank other models by similarity • Measure how many models in the same class appear near the top of the ranked list • Present average results Query

10. Typical Retrieval Experiment • Create a database of 3D models • Group the models into classes • For each model: • Rank other models by similarity • Measure how many models in the same class appear near the top of the ranked list • Present average results Query

11. Typical Retrieval Experiment • Create a database of 3D models • Group the models into classes • For each model: • Rank other models by similarity • Measure how many models in the same class appear near the top of the ranked list • Present average results

12. Typical Retrieval Experiment • Create a database of 3D models • Group the models into classes • For each model: • Rank other models by similarity • Measure how many models in the same class appear near the top of the ranked list • Present average results

13. Shape Retrieval Results

14. Outline • Introduction • Related work • Princeton Shape Benchmark • Comparison of 12 descriptors • Evaluation techniques • Results • Conclusion

15. Typical Shape Databases

16. Typical Shape Databases

17. Aerodynamic Typical Shape Databases

18. Letter ‘C’ Typical Shape Databases

19. Typical Shape Databases

20. Typical Shape Databases

21. Typical Shape Databases 153 dining chairs 25 living room chairs 16 beds 12 dining tables 8 chests 28 bottles 39 vases 36 end tables

22. Typical Shape Databases

23. Goal: Benchmark for 3D Shape Retrieval • Large number of classified models • Wide variety of class types • Not too many or too few models in each class • Standardized evaluation tools • Ability to investigate properties of descriptors • Freely available to researchers

24. Princeton Shape Benchmark • Large shape database • 6,670 models • 1,814 classified models, 161 classes • Separate training and test sets • Standardized suite of tests • Multiple classifications • Targeted sets of queries • Standardized evaluation tools • Visualization software • Quantitative metrics

25. 51 potted plants 33 faces 15 desk chairs 22 dining chairs 100 humans 28 biplanes 14 flying birds 11 ships Princeton Shape Benchmark

26. Princeton Shape Benchmark (PSB)

27. Princeton Shape Benchmark (PSB)

28. Outline • Introduction • Related work • Princeton Shape Benchmark • Comparison of 12 descriptors • Evaluation techniques • Results • Conclusion

29. Comparison of Shape Descriptors • Shape Histograms (Shells) • Shape Histograms (Sectors) • Shape Histograms (SecShells) • D2 Shape Distributions • Extended Gaussian Image (EGI) • Complex Extended Gaussian Image (CEGI) • Spherical Extent Function (EXT) • Radialized Spherical Extent Function (REXT) • Voxel • Gaussian Euclidean Distance Transform (GEDT) • Spherical Harmonic Descriptor (SHD) • Light Field Descriptor (LFD)

30. Comparison of Shape Descriptors

31. Evaluation Tools Visualization tools • Precision/recall plot • Best matches • Distance image • Tier image Quantitative metrics • Nearest neighbor • First and Second tier • E-Measure • Discounted Cumulative Gain (DCG)

32. Evaluation Tools Visualization tools • Precision/recall plot • Best matches • Distance image • Tier image Quantitative metrics • Nearest neighbor • First and Second tier • E-Measure • Discounted Cumulative Gain (DCG)

33. Evaluation Tools Query Correct class Visualization tools • Precision/recall plot • Best matches • Distance image • Tier image Quantitative metrics • Nearest neighbor • First and Second tier • E-Measure • Discounted Cumulative Gain (DCG) Wrong class

34. Evaluation Tools Visualization tools • Precision/recall plot • Best matches • Distance image • Tier image Quantitative metrics • Nearest neighbor • First and Second tier • E-Measure • Discounted Cumulative Gain (DCG)

35. Evaluation Tools Visualization tools • Precision/recall plot • Best matches • Distance image • Tier image Quantitative metrics • Nearest neighbor • First and Second tier • E-Measure • Discounted Cumulative Gain (DCG)

36. Evaluation Tools Visualization tools • Precision/recall plot • Best matches • Distance image • Tier image Quantitative metrics • Nearest neighbor • First and Second tier • E-Measure • Discounted Cumulative Gain (DCG) Dining ChairDesk Chair

37. Rectangular table Function vs. Shape • Functional at the top levels of the hierarchy, shape based at the lower levels root Man-made Natural Vehicle Furniture Table Chair Round table

38. Base Classification (92 classes) Man-made Furniture Table Round table

39. Coarse Classification (44 classes) Man-made Furniture Table Round table

40. Coarser Classification (6 classes) Man-made Furniture Table Round table

41. Coarsest Classification (2 classes) Man-made Furniture Table Round table

42. Granularity Comparison Base(92) Man-made vs. Natural (2)

43. Rotationally Aligned Models (650)

44. All Models (907)

45. Complex Models (200)

46. Performance by Property

47. Conclusion • Methodology to compare shape descriptors • Vary classifications • Query lists targeted at specific properties • Unexpected results • EGI: good at discriminating man-made vs. natural objects, though poor at fine-grained distinctions • LFD: good overall performance across tests • Freely available Princeton Shape Benchmark • 1,814 classified polygonal models • Source code for evaluation tools

48. Future Work • Multi-classifiers • Evaluate statistical significance of results • Application of techniques to other domains • Text retrieval • Image retrieval • Protein classification

49. Acknowledgements David Bengali partitioned thousands of models. Ming Ouhyoung and his students provided the light field descriptor. Dejan Vranic provided the CCCC and MPEG-7 databases. Viewpoint Data Labs donated the Viewpoint database. Remco Veltkamp and Hans Tangelder provided the Utrecht database. Funding: The National Science Foundation grants CCR-0093343 and 11S-0121446.

50. The End http://shape.cs.princeton.edu/benchmark