Two approaches to robust hand pose estimation : generative modeling and semantic relations

Abstract

This thesis explores hand pose estimation through the use of two methods. We first investigate the use of segmentation networks within pose estimation pipelines with a focus on fine parts segmentation. We present two implementations of a novel method for fine parts segmentation employing a higher-order Conditional Random Field (CRF) that measures attachment and containment of fine parts. The first implementation is of the CRF as a post-processing module on top of a Convolutional Neural Network (CNN). The second addresses efficiency bottlenecks in the first by implementing the CRF as a Recurrent Neural Network (RNN) and allowing for end-to-end training with the CNN. Limited by the accuracy of fine parts segmentation and wishing to avoid propagation of segmentation errors through a pipeline, we turn to generative modeling methods for hand pose estimation and present an inverse-graphics approach implemented in a probabilistic programming language. Spurred by the lack of occlusion in hand image datasets, we present the MIT Partially Occluded Hands Dataset, a large-scale dataset of single RGB images, half of which feature natural hand-object interactions, and evaluate several baselines on this dataset.