Detailed Human Shape and Pose from Images

Abstract

Much of the research on video-based human motion capture assumes the body shape is known

Shining a Light on Human Pose: On Shadows, Shading and the Estimation of Pose and Shape

“…To estimate the model and light parameters, we formulate an objective function in terms of a silhouette dissimilarity measure, D(· e ,· o ), which is implemented as a bidirectional Chamfer distance between estimated and observed silhouettes [3]. In [3] we rely on foreground silhouettes alone to estimate pose and shape; here we add a term to measure the shadow difference.…”

“…In [3] we relied on a cylindrical body model to perform human tracking and initialize the SCAPE optimization. This was shown to work well in a multi-camera setup, but still required initialization at the first frame.…”

“…3. Estimate pose and shape parameters from foreground silhouette data alone and generate the surface meshes in each frame (Section 2.3) [3]. 4.…”

“…In recent work [3] we used a detailed graphics body model (SCAPE) [1], learned from range scans of real people, to address the problem of markerless human pose and shape estimation in a multicamera setting ( Figure 2). The generative model predicts silhouettes in each camera view given the pose/shape parameters of the the body and matches them to foreground silhouettes extracted from images using a fairly standard Chamfer distance measure.…”

“…For a complete framework, we must also estimate the lighting in the scene automatically. We propose two approaches that exploit 3D body pose and shape represented using the SCAPE model [1], the parameters of which are estimated directly from image foreground silhouettes (without knowledge of scene illumination) [3].…”