SiamCorners: Siamese Corner Networks for Visual Tracking

The current Siamese network based on region proposal (RPN) has attracted great attention in visual tracking due to its excellent accuracy and high efficiency. However, the design of RPN involves selection number, scale, aspect ratios anchor boxes, which will affect applicability convenience model. Furthermore, these boxes require complicated calculations, such as calculating their intersection-over-union (IoU) with ground truth bounding boxes. Due problems related we propose a simple yet effective anchor-free tracker (named corner networks, SiamCorners), is end-to-end trained offline large-scale image pairs. Specifically, introduce modified pooling layer convert box estimate target into pair predictions (the bottom-right top-left corners). By corners, avoid need This make entire algorithm more flexible than anchor-based trackers. In our design, further layer-wise feature aggregation strategy that enables module predict multiple corners for deep networks. We then new penalty term used select an optimal candidate corners. Finally, SiamCorners achieves experimental results are comparable state-of-art while maintaining running speed. particular, 53.7% AUC NFS30 61.4% UAV123, still at 42 frames per second (FPS).