Spatio-temporal matching for siamese visual tracking

Siamese trackers formulate the visual tracking task as a similarity matching problem through cross correlation. It is arduous for such methods to track targets with presence of distractors. We suspect reasons are twofold: 1) The irrelevant activated channels in correlation map will produce ambiguous results. 2) pipeline per-frame process and cannot handle response aberrance caused by temporal context variation. In this paper, we propose spatio-temporal thoroughly explore capability 4-D space (height, width channel) time. spatial matching, introduce space-variant instance-aware (SI-Corr) implement different channel-wise recalibration each position. SI-Corr can guide generation features distinguish target distractors at instance level. design an repressed module (ARM) investigate short-term positional relationship between ARM utilizes simple optimization method restrict abrupt alteration interframe maps, which allows network learn consistency structure distribution. Moreover, efficiently embed into inference process. Experiments on six benchmarks, including OTB100, VOT2018, VOT2020, GOT-10k, LaSOT TrackingNet, demonstrate state-of-the-art performance proposed method.