Fully Convolutional Online Tracking

Online learning has turned out to be effective for improving tracking performance. However, it could be simply applied for classification branch, but still remains challenging for adapting to regression branch due to the complex design. To tackle this issue, we present the first fully convolutional online tracking framework (FCOT), with a focus on enabling online learning for both classification and regression branches. Our key contribution is to introduce an online regression model generator (RMG) based on the carefully designed anchor-free box regression branch, which enables our FCOT to be more effective in handling target deformation during tracking procedure. In addition, to deal with the confusion of similar objects, we devise a simple yet effective multi-scale classification branch to improve both accuracy and robustness of FCOT. Due to its simplicity in design, our FCOT could be trained and deployed in a fully convolutional manner with a running speed of 45FPS. The proposed FCOT sets a new state-of-the-art results on six benchmarks including VOT2018, LaSOT, TrackingNet, GOT-10k, UAV123, and NFS. Particularly, among real-time trackers, our FCOT achieves the best EAO of 0.508 on VOT2018, NP of 0.678 on LaSOT, NP of 0.828 on TrackingNet, and AO of 0.640 on GOT-10k. The code and models will be made available at https://github.com/MCG-NJU/FCOT.