Self-Paced Dense Connectivity Learning for Visual Tracking

When misalignment, deformation, and tracking failures occur, the appearance of the target tends to change significantly. How to effectively learn the change of target's appearance is an essential problem in visual tracking. Recently, most recent trackers based on convolutional neural networks update the tracker online to learn the change of target's appearance. These methods collect tracking results as online training samples. Thus, the reliability of training samples is very important for online updates. We propose a self-paced selection model, which integrates the self-paced learning model into the tracking framework for the goal of distinguishing the reliable samples from the tracking results. It estimates the reliability of the tracking results by the self-paced function. We design a method that adaptively calculates the value of the pace, which determines the number of samples selected. And this method is based on the number of tracking results. At the same time, the quality of the target's features plays a key role in the performance of the tracker. We employ dense connectivity learning to enhance the flow of information throughout the network, which makes the target's features represent better. The extensive experiments demonstrate that our self-paced dense connectivity learning tracker (SPDCT) performs favorably against the state-of-the-art trackers over four benchmark datasets.