Shape from Video: Dense Shape, Texture, Motion and Lighting from Monocular Image Streams
Azeem Lakdawalla and Aaron Hertzmann
University of Toronto
Abstract: This paper presents a probabilistic framework for robust
recovery of dense 3D shape, motion, texture and lighting
from monocular image streams. We assume that the
object is smooth, Lambertian, illuminated by one distant
light source, and subject to smoothly-varying rigid motion.
The problem is formulated as a MAP estimation problem in
which all shape, motion, noise variances and outlier probabilities
are estimated simultaneously. Estimation is performed
using a multi-stage initialization process followed
by a large-scale quasi-Newtonian optimization technique.
A. Lakdawalla and A. Hertzmann. Shape From Video: Dense Shape, Texture, Motion and Lighting from Monocular Image Streams. IEEE Workshop on Photometric Analysis for Computer Vision, in conjunction with ICCV '07, Rio de Janeiro, Brazil.
Results
| Input Sequence | Recovered Shape |
|---|
 |  |  |
 |  |  |
 |  |  |
| Occluded Sequence | Using our robust method | Without robustness |
|---|
 |  |  |  |  |