Estimating Human Shape and Pose from a Single

Estimating Human Shape and Pose from a Single Image Peng Guan Alex Weiss Alexandru Balan Michael J. Black Brown University Department of Computer Science ICCV’ 2009

Body shape and pose from 1 image?

Introduction What others do • Estimating 3 D human pose in uncalibrated monocular imagery • Use silhouette in multicamera setting to recover 3 D body shape • Most work assumes the existence of a known background to extract foreground silhouette • In previous body models, height is correlated with other shape variations What we do • Estimating both 3 D shape and pose in uncalibrated monocular imagery • Use additional monocular cues including smooth shading • Use Grab. Cut to produce foreground region • Make height variation concentrated along one shape basis vector, which allows “height constrained fitting”

Previous Work 3 D pose and shape estimation from multiple, calibrated, cameras Balan, A. , Sigal, L. , Black, M. J. , Davis, J. , Haussecker, H, “Detailed human shape and pose from images”, Proc. IEEE Conf. on Computer Vision and Pattern Recognition, CVPR, Minneapolis, June 2007

SCAPE Body Model D. Anguelov, P. Srinivasan, D. Koller, S. Thrun, J. Rodgers, and J. Davis. SCAPE: Shape completion and animation of people. SIGGRAPH, 24(3): 408– 416, 2005.

Body shape/pose from 1 image: Problems 1. High dimensional body model (shape and pose) – initialization problem. 2. Background unknown 3. Single, monocular image 1. poorly constrained 2. Shape/Pose ambiguities 4. Silhouette insufficient

Solution 1: Pose Initialization Better Shape: initialized to mean body shape.

Solution 2: Segmentation C. Rother, V. Kolmogorov, and A. Blake. “Grab. Cut”: Interactive foreground extraction using iterated graph cuts. SIGGRAPH, 23(3): 309– 314, 2004.

Problem: Pose/Shape ambiguities Body shape and pose fitted to a single camera view

Solution 3: Height Preserving Shape Space

Shape space without height preserving

Problem: Silhouette not sufficient

Solution 4: Edge Cues

Problem: Shape not well constrained

Solution 5: Parametric Shape from Shading M. de la Gorce, N. Paragios and David Fleet. Model-Based Hand Tracking with Texture, Shading and Selfocclusions. IEEE Conference in Computer Vision and Pattern Recognition (CVPR), Anchorage 2008.

Shading/Overall Cost function Shading cost function: Overall cost function:

Experiment: Lab Images

Experiment: Lab Images

Quantitative Comparison

Experiment: Internet Images

Experiment: Paintings

Conclusions Contributions • Solution to a new problem: Human pose and shape estimation from a single image • Parametric shape from shading for estimating human shape from complex images and paintings • Attribute-constrained body model Limitations • Single point light assumption and simplified model of surface reflection • User assistance for pose initialization • Minimal clothing for shading

Acknowledgement • Financial support: NSF IIS-0812364 and the RI Economic Development Corp. • Peng Guan, Alexander Weiss, Alexandru Balan, Michael Black, “Estimating Human Shape and Pose from a Single Image”, Int. Conf. on Computer Vision, ICCV, Kyoto, Japan, Sept. 2009 • Alexander Weiss: Grab. Cut 3 D pose initialization • Alexandru Balan: Height preserving shape space • David Hirshberg: Projection of model edge

Thank you!