Improving gesture recognition through spatial focus of attention
Date
2018
Authors
Narayana, Pradyumna, author
Draper, Bruce A., advisor
Beveridge, Ross J., committee member
Anderson, Charles W., committee member
Peterson, Christopher, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
Gestures are a common form of human communication and important for human computer interfaces (HCI). Most recent approaches to gesture recognition use deep learning within multi- channel architectures. We show that when spatial attention is focused on the hands, gesture recognition improves significantly, particularly when the channels are fused using a sparse network. We propose an architecture (FOANet) that divides processing among four modalities (RGB, depth, RGB flow, and depth flow), and three spatial focus of attention regions (global, left hand, and right hand). The resulting 12 channels are fused using sparse networks. This architecture improves performance on the ChaLearn IsoGD dataset from a previous best of 67.71% to 82.07%, and on the NVIDIA dynamic hand gesture dataset from 83.8% to 91.28%. We extend FOANet to perform gesture recognition on continuous streams of data. We show that the best temporal fusion strategies for multi-channel networks depends on the modality (RGB vs depth vs flow field) and target (global vs left hand vs right hand) of the channel. The extended architecture achieves optimum performance using Gaussian Pooling for global channels, LSTMs for focused (left hand or right hand) flow field channels, and late Pooling for focused RGB and depth channels. The resulting system achieves a mean Jaccard Index of 0.7740 compared to the previous best result of 0.6103 on the ChaLearn ConGD dataset without first pre-segmenting the videos into single gesture clips. Human vision has α and β channels for processing different modalities in addition to spatial attention similar to FOANet. However, unlike FOANet, attention is not implemented through separate neural channels. Instead, attention is implemented through top-down excitation of neurons corresponding to specific spatial locations within the α and β channels. Motivated by the covert attention in human vision, we propose a new architecture called CANet (Covert Attention Net), that merges spatial attention channels while preserving the concept of attention. The focus layers of CANet allows it to focus attention on hands without having dedicated attention channels. CANet outperforms FOANet by achieving an accuracy of 84.79% on ChaLearn IsoGD dataset while being efficient (≈35% of FOANet parameters and ≈70% of FOANet operations). In addition to producing state-of-the-art results on multiple gesture recognition datasets, this thesis also tries to understand the behavior of multi-channel networks (a la FOANet). Multi- channel architectures are becoming increasingly common, setting the state of the art for performance in gesture recognition and other domains. Unfortunately, we lack a clear explanation of why multi-channel architectures outperform single channel ones. This thesis considers two hypotheses. The Bagging hypothesis says that multi-channel architectures succeed because they average the result of multiple unbiased weak estimators in the form of different channels. The Society of Experts (SoE) hypothesis suggests that multi-channel architectures succeed because the channels differentiate themselves, developing expertise with regard to different aspects of the data. Fusion layers then get to combine complementary information. This thesis presents two sets of experiments to distinguish between these hypotheses and both sets of experiments support the SoE hypothesis, suggesting multi-channel architectures succeed because their channels become specialized. Finally we demonstrate the practical impact of the gesture recognition techniques discussed in this thesis in the context of a sophisticated human computer interaction system. We developed a prototype system with a limited form of peer-to-peer communication in the context of blocks world. The prototype allows the users to communicate with the avatar using gestures and speech and make the avatar build virtual block structures.
Description
Rights Access
Subject
focus of attention
network fusion
gesture recognition
deep learning