Examination Committee

Prof Yiwen WANG, ECE/HKUST (Chairperson)
Prof Ling SHI, ECE/HKUST (Thesis Supervisor)
Prof Lu FANG, ECE/HKUST

Abstract

In this thesis work, we propose several effective techniques to train deep convolution neural networks(CNN) for hand pose estimation, which is a complex regression problem. We propose repeating-ladder-type learning rate to avoid poor saddle points and hierarchical training based on physical property grouping to improve the final prediction accuracy. These two techniques can migrate to other regression problems which share similar properties with the hand pose estimation problem. We also present several image pre-processing  methods to improve the stability of the training process. During the training process, we use high quality synthesized images to pre-train the networks and then use noisy initial images for fine-tuning. We test on the NYU Hand Pose Dataset. Our method achieves the state-of-the-art result in the middle precision region and near the state-of-the-art result in the high precision region.