High-speed High-performance Visual Tracker via Correlation Filter with Compressed Deep Feature

Abstract

This paper introduces a context-aware correlation filter based tracker to achieve both high speed and high performance. We achieve high speed via deep feature compression based on a context-aware scheme utilizing multiple expert auto-encoders. To achieve high performance with the compressed feature map, we introduce extrinsic denoising processes and a new orthogonality loss term for pre-training and fine-tuning of the expert auto-encoders. In experiments, the proposed tracker is verified to achieve a comparable performance to state-of-the-art with running at over 100 fps.