Pyramid Correlation based Deep Hough Voting for Visual Object Tracking

• DOI: 10.48550/arxiv.2110.07994
• Publication Date: 2021-01-01
• Authors: Ying Wang; Tingfa Xu; Jianan Li; Shenwang Jiang; Junjie Chen
• Keywords: Pyramid (geometry); Feature (linguistics); BitTorrent tracker

Most of the existing Siamese-based trackers treat tracking problem as a parallel task classification and regression. However, some studies show that sibling head structure could lead to suboptimal solutions during network training. Through experiments we find that, without regression, performance be equally promising long delicately design suit training objective. We introduce novel voting-based classification-only algorithm named Pyramid Correlation based Deep Hough Voting (short for PCDHV), jointly locate top-left bottom-right corners target. Specifically innovatively construct module equip embedded feature with fine-grained local structures global spatial contexts; The elaborately designed further take over, integrating long-range dependencies pixels perceive corners; In addition, prevalent discretization gap is simply yet effectively alleviated by increasing resolution maps while exploiting channel-space relationships. general, robust simple. demonstrate effectiveness through series ablation experiments. Without bells whistles, our tracker achieves better or comparable SOTA algorithms on three challenging benchmarks (TrackingNet, GOT-10k LaSOT) running at real-time speed 80 FPS. Codes models will released.