430 likes | 594 Views
A Photometric Approach for Estimating Normals and Tangents. Input: Images under varying point lighting. Output: Estimate of surface orientation. Normal Field. Tangent Field. Related Work Lambertian photometric stereo [Woodham 1980].
E N D
Output: Estimate of surface orientation Normal Field Tangent Field
Related Work • Lambertian photometric stereo [Woodham 1980] • Discard specular highlight [Coleman and Jain 1982; Mallick et al. 2005] [Klette et al. 1998]
Fit to low-dimensional parametric models [Georghiades 2003; Goldman et al. 2005] • Fit non-parametric curves (isotropic only) [Alldrin et al. 2008]
Locate mirror direction [Wang and Dana 2006; Chen et al. 2006; Ma et al. 2007; Francken et al. 2008; Nehab et al. 2008] Image: Ma et al. 2007
Our Approach 2D slice of the BRDF (fixed view) Half-angle parameterization
Our Approach 2D slice of the BRDF (fixed view) Half-angle parameterization
Analysis of microfacet-based models: • Microfacet distributionAlmost all analytic and measured distributions exhibit these symmetries. Images: Ngan et al. 2005
Analysis of microfacet-based models: • Fresnel TermWell approximated by , and is asymmetric only at grazing angles. 1 -90° 90°
Analysis of microfacet-based models: • Shadowing/MaskingSmooth and can be greatly simplified. Shadowing/Masking function from Wang et al. 2008 • [Torrance1987; Ashikhmin et al. 2000;Ngan et al. 2005; Wang et al. 2008]
Restriction of light positions view view normal
Restriction of light positions view normal
Restriction of light positions view normal
Validation Normal Error Tangent Error BRDF from Ngan et al. 2005
Validation Normal Error Tangent Error BRDF from Ngan et al. 2005
Anisotropic Ward Measured Purple Satin Yellow Satin Brushed Metal Ngan et al. 2005
Error Analysis (Torrance-Sparrow) Our approach Photometric stereo Specularity stereo Normal Error (Degrees) Diffuse Shiny
Acquisition Calibrated spherical gantry 1,500 1024x1024 HDR images 2.3 GB 45 minutes
Algorithm 1. Reconstruct a continuous 2D slice of the BRDF at each pixel using barycentric interpolation of the original data. 2. Estimate by optimizing 3. Estimate by holding fixed and optimizing
Implementation Independent at each pixel 42x Dual 1.6 Ghz Opertons 10 minutes (~7 hours for single machine)
Limitations Interreflections
Limitations: Sampling density 1,512 743 380 172
Limitations Non-symmetric microfacet distributions Red velvet dataset from Ngan et al. [2005]
Conclusion • Main advantages: • General, does not rely on parametric model • First technique to directly recover tangent field
Thank You Acknowledgments • Jiajun Zhu for help with data capture • NSF CAREER award CCF-0747220 • NSF grant CCF-0811493 • NVIDIA Professor Partnership award • Fellowship from the Sloan Foundation