Modeling Hair from Multiple Views

1. Modeling Hair from Multiple Views Y. Wei, E. Ofek, L. Quan and H. Shum * Hong Kong University of Science and Technology HK UST

2. Hair is a defining characteristic of human appearance

3. Overview • Related work • Principle of our approach • Implementation • Experimental results • Conclusion

4. Related work • Manual modeling Hadap and Magnenat-Thalmann 00, Kim et al. 02 (-) Tedious user interaction

5. Related work • Image based rendering Matusik et al. 02 (+) Photo-Realistic (-) Long capture, Large storage (-) No animation

6. Related work Paris et al. 04, Grabli et al. 02 • Image based modeling • Difficult due to insufficient image resolution • Typically 500 pixels for 50,000 hair fibers • Hair strands, groups of hair fibers, are visible

7. Paris et al. 2004 • Hundreds of images • Fixed view point • Controlled and variable illuminations (+) Impressive results (-) Complex capturing (-) Controlled illumination (-) limited visibility

8. Our goals Modeling from multiple view points • Easy capturing • Complex hair styles • Enable capturing dynamic scenes

9. Image Capturing • Hand-Held camera • Natural illumination condition

10. 2D orientation map 2D image orientation of hair strands • Using oriented filter as in Paris’ work

11. 3D direction in one view 2D direction constrains the growth direction to a plane in space. Hair segment

12. N D=NxN’ N’ D 3D direction in two views • Two images define a unique 3D direction

13. D minimizes Σj (Nj · D)2 σj2 Where ║D║ = 1 Nj 2D orientation confidence 3D direction in multi-views • Each image pair defines a unique direction • Multiple images – Linear optimization D

14. 3D Hair Growth according to 2D orientation Hair fiber – a sequence of chained segments Hair segment Initial growth direction – outward from scalp

15. Implementation • Camera calibration • Hair volume definition and fiber initialization • Fiber growth and visibility determination • Growth termination

16. Camera calibration Automatic calibration – quasi-dense approach (Lhuillier and Quan 05)

17. Hair volume • Shair , hair surface • Estimated from images Shair

18. Approximating Shair using visual hull

19. Hair volume • Sscalp, scalp surface • Generic head model Sscalp • Ellipsoid • Offset of Shair

20. Hair initialization • Generate root points on Sscalp • Define hair coverage of Sscalp by projection on hair masks

21. Fiber Visibility • Visible P – visibility determined by hair surface geometry P

22. Fiber Visibility • P internal: Visibility is determined by the closest surface point. P’ P Shair

23. Fiber Visibility • P internal: The deeper P is, the smoother the growth direction.

24. Growth Termination Shair Visual hull Sscalp • Exceeds pre-defined length • Grows outside the hair bounding surface • Large inconsistency between views

25. An Example of Hair Growth

26. Experiments : Dark Hair

27. Smooth hair

29. Curly hair

30. Conclusions • Multi view approach • Simple capture • Open possibilities for hair motion capture. • Automatic process (masks are still manual). • High quality results.

31. Future work • Capturing of dynamic hair using a set of multiple cameras • Recovery of hair appearance

32. Thanks Previous work appear by courtesy of Sylvain Paris, Tae-Yong Kim and Wojciech Matusik. Some rendering provided by Florence Bertails, Hubert Nguyen

33. Thanks Previous work appear by courtesy of Sylvain Paris, Tae-Yong Kim and Wojciech Matusik. Some rendering provided by Florence Bertails, Hubert Nguyen

34. Thank You! HK UST