Siamese Visual Tracking With Residual Fusion Learning

Multi-stage feature fusion is pretty effective for deep Siamese trackers to promote tracking performance. Unfortunately, conventional approaches, such as weighted average, are so simple that they inappropriate combine the features with diverse characteristics. In addition, module generally optimized along network module, which may result in performance degradation of whole tracker. this paper, we propose a novel tracker by exploiting expression capacity residual learning (SiamRFL). Specifically, employs deep-layer direct input semantically recognize object from background, and refines state local detail patterns exploring shallow-layer through channel. The classification regression can be fused respectively deploying multiple units. To avoid problem, also present an ensemble training framework our tracker, different loss functions introduced individually optimize modules. Compared baseline SiamRPN++ proposed achieves favorable gains $0.696\rightarrow 0.709$ , notation="LaTeX">$0.285\rightarrow 0.308$ notation="LaTeX">$0.603\rightarrow 0.624$ notation="LaTeX">$0.496\rightarrow 0.520$ notation="LaTeX">$0.517\rightarrow 0.559$ on OTB100, VOT2019, UAV123, LaSOT GOT10k datasets, outperforming other approaches obvious margin.