HROM: Learning high-resolution representation and object-aware masks for visual object tracking

Abstract

Siamese network-based trackers consider tracking as features cross-correlation between the target template and the search region. Therefore, feature representation plays an important role for constructing a high-performance tracker. However, all existing Siamese networks extract the deep but low-resolution features of the entire patch, which is not robust enough to estimate the target bounding box accurately. In this work, to address this issue, we propose a novel high-resolution Siamese network, which connects the high-to-low resolution convolution streams in parallel as well as repeatedly exchanges the information across resolutions to maintain high-resolution representations. The resulting representation is semantically richer and spatially more precise by a simple yet effective multi-scale feature fusion strategy. Moreover, we exploit attention mechanisms to learn object-aware masks for adaptive feature refinement, and use deformable convolution to handle complex geometric transformations. This makes the target more discriminative against distractors and background. Without bells and whistles, extensive experiments on popular tracking benchmarks containing OTB100, UAV123, VOT2018 and LaSOT demonstrate that the proposed tracker achieves state-of-the-art performance and runs in real time, confirming its efficiency and effectiveness.