Efficient codes for multi-modal pose regression


Redundancy reduction, or sparsity, appears to be an important information-theoretic principle for encoding natural sensory data. While sparse codes have been the subject of much recent research, they have primarily been evaluated using readily available datasets of natural images and sounds. In comparison, relatively little work has investigated the use of sparse codes for representing information about human movements and poses. This paper proposes a basic architecture for evaluating the impact of sparsity when coding human poses, and tests the performance of several coding methods within this framework for the task of mapping from a kinematic (joint angle) modality to a dynamic (joint torque) one. We show that sparse codes are indeed useful for effective mappings between modalities and examine in detail the sources of error for each stage in the model.

Back to Table of Contents