Ocean: Object-aware Anchor-free Tracking with Matching-relation Learning

Abstract Improving the tolerance for accumulated errors in object localization and enhancing the stability for estimating matching relations are key issues in robust Siamese network-based single object tracking in real-time. In this paper, we propose an object-aware anchor-free network for object tracking. Different from refining the reference anchors, the position and scale of the object are predicted in an anchor-free way. Since each position within a ground-truth box is trained, inaccurate predictions of the object can be rectified during tracking. An irregular sampling-based alignment module is designed to extract object-aware features from the predicted bounding box. The object-aware features are used to improve the classification of the predicted bounding box into the object or background. In order to improve the stability of matching-relation learning for object tracking, we upgrade the Siamese network by using automated search of the matching network based on binary channel operations. Not relying on explicit similarity calculation, by combining relation-operators, different matching networks are searched for the classification and regression tasks for tracking. The adaptability of relation-learning to different tasks is enhanced. Experimental results on many tracking benchmarks show the effectiveness of the proposed trackers.