A robust likelihood function for 3D human pose tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Unsupervised 3D human pose tracking often relies on optimizing the likelihood function.
• Existing methods primarily focus on improving the optimization framework (tracker).

Purpose of the Study:

• To enhance unsupervised 3D human pose tracking by improving the likelihood function.
• To make the optimization task easier by creating a more robust and less ambiguous likelihood function.

Main Methods:

• Proposed an exponential chamfer distance for robust model matching against small pose variations.
• Introduced a part-based model for improved localization of occluded and overlapping body parts.
• Utilized a standard annealing particle filter and a simple diffusion motion model.

Main Results:

• Achieved significantly lower tracking error compared to other unsupervised methods on the HumanEva dataset.
• Developed a novel bias correction method to transform between different joint systems.
• Part-based likelihood with bias correction yielded results comparable to state-of-the-art supervised tracking methods.

Conclusions:

• Improving the likelihood function is a viable and effective strategy for unsupervised 3D human pose tracking.
• The proposed methods enhance robustness and localization accuracy, even with partial occlusions.
• Bias correction enables unsupervised methods to reach supervised state-of-the-art performance levels.