Efficient face recognition for wearable camera devices
Titan Sécurité Inc. has deployed wearable video camera devices for security and surveillance applications, and seeks to accurately detect and recognize objects appearing in captured videos. This project focuses on video-based face recognition (FR), where facial trajectories captured with video cameras are compare against one (or few) reference stills for each individual of interest. The performance of these FR systems is typically poor due to complex and changing video surveillance environments, e.g., variations of facial appearance due to pose, illumination, blur, etc. Given the state-of-the-art accuracy achieved with deep learning architectures on many challenging visual recognition problems, Titan Sécurité Inc. seeks to design Siamese networks based on deep convolutional neural networks (CNNs) for still-to-video FR. However, since these networks represent complex solutions for real-time applications, this project seeks to develop specialized techniques to reduce their time and memory complexity. These include advanced techniques for reducing search time, selecting features, and pruning parameter. In particular, this project will focus on developing filter-level pruning techniques that can simultaneously accelerate and compress a CNN based on information extracted from its layers.