Recently, deep trackers based on the siamese networking are enjoying increasing popularity in the tracking community. Generally, those trackers learn a high-level semantic embedding space for feature representation but lose low-level fine-grained details. Meanwhile, the learned high-level semantic features are not updated during online tracking, which results in tracking drift in presence of target appearance variation and similar distractors. In this paper, we present a novel end-to-end trainable Convolutional Neural Network (CNN) based on the siamese network for distractor-aware tracking. It enhances target appearance representation in both the offline training stage and online tracking stage. In the offline training stage, this network learns both the low-level fine-grained details and high-level coarse-grained semantics simultaneously in a multi-task learning framework. The low-level features with better resolution are complementary to semantic features and able to distinguish the foreground target from background distractors. In the online stage, the learned low-level features are fed into a correlation filter layer and updated in an interpolated manner to encode target appearance variation adaptively. The learned high-level features are fed into a cross-correlation layer without online update. Therefore, the proposed tracker benefits from both the adaptability of the fine-grained correlation filter and the generalization capability of the semantic embedding. Extensive experiments are conducted on the public OTB100 and UAV123 benchmark datasets. Our tracker achieves state-of-the-art performance while running with a real-time frame-rate.
Fine-Grained Representation Learning and Recognition by Exploiting Hierarchical Semantic Embedding