SiamCA: Siamese visual tracking with customized anchor and target-aware interaction

Abstract

Generating high-quality candidates is very crucial for discriminative target tracking. Siamese network based trackers rely on some fixed anchors to fit the shape of a target. However, in practical tracking, these fixed anchors are insufficient to cover all possible aspect ratios of the target caused by various motions and deformations. In this paper, we propose customized anchor that can fit the current state of the target using only one bounding box at each coordinate point. The customized anchor is generated under the guidance of historical prediction result and initial target state, and is further refined by embedding its shape feature into the regression map. To provide more discriminative feature for generating customized anchor, we design a target-aware based feature correlation module to capture local spatial context and global interaction of target, by which the target feature is more salient while background interference is effectively suppressed. In addition, we propose a dynamic template update method and a simple yet effective lost-target re-search strategy to find the lost-target again. Experiments on benchmarks including OTB100, VOT2019, LaSOT, UAV123, and VOT2018 demonstrate that our tracker achieves promising performance with a real-time speed. Our code and pre-trained models are available at https://github.com/PPs199/SiamCA.git.