Multi-scale Region Proposal Network Trained by Multi-domain Learning for Visual Object Tracking

초록

This paper presents a multi-scale region proposal network (RPN) for visual object tracking, inspired by Faster R-CNN and Yolo detectors which adopt an RPN to significantly speed up the detection time and achieve state-of-the-art detection performance. We expand them to apply a multi-scale region proposal network for visual tracking. Our proposed network can utilize both fine-grained features from shallow convolutional layers and discriminative features from deep convolutional layers. The features of shallow layers are good at accurate objects localization, and the features of deep convolutional layers can efficiently distinguish between target objects and backgrounds. A multi-domain learning mechanism is applied to train our network in an end-to-end way. To predict a new target object and its location in a new frame, we propose an re-ranking algorithm to determine a true object by exploiting spatial modeling, scale variants and color attributes of object proposals. Our tracker is validated on the OTB-15 object tracking benchmark, and achieves 0.603 for the success rate and 0.760 for the precision rate of the one-pass evaluation. Additionally, our tracker can run at 22 frames per second, which is very close to real-time speed. Experiment results show its outstanding performance in both tracking accuracy and speed by comparing it with existing state-of-the-art methods.