Recovering 3D human pose from monocular images.

Area of Science:

• Computer Vision
• Machine Learning
• Human Pose Estimation

Background:

• Recovering 3D human body pose from 2D images is challenging due to depth ambiguity and lack of explicit body models.
• Existing methods often rely on detailed prior information like body part labeling or explicit 3D models.

Purpose of the Study:

• To develop a robust, learning-based method for 3D human body pose recovery from single images and monocular sequences.
• To address the inherent ambiguity in 3D pose recovery from silhouettes without explicit body models or part labels.

Main Methods:

• Direct nonlinear regression using histogram-of-shape-contexts descriptors from image silhouettes.
• Evaluation of regression techniques including Ridge, Relevance Vector Machine (RVM), and Support Vector Machine (SVM) regression.
• Integration into a regressive tracking framework to resolve pose ambiguities using temporal dynamics.

Main Results:

• Relevance Vector Machines (RVMs) offered sparser regressors with comparable performance to other methods.
• Kernel bases provided a slight but significant performance improvement.
• The regressive tracking framework demonstrated stable tracking of long image sequences.

Conclusions:

• The proposed method effectively recovers 3D human body pose from monocular images without explicit body models or part labeling.
• The histogram-of-shape-contexts and RVM regression, combined with a novel tracking framework, provide robust and accurate pose estimation.
• The system achieved mean angular errors of 4-6 degrees for walking motions, demonstrating its effectiveness on real-world data.