Fully convolutional online tracking

Online learning has turned out to be effective for improving tracking performance. However, it could simply applied classification branch, but still remains challenging adapt regression branch due its complex design and intrinsic requirement high-quality online samples. To tackle this issue, we present the fully convolutional framework, coined as FCOT, focus on enabling both branches by using a target filter based paradigm. Our key contribution is introduce an model generator (RMG) initializing weights of in with samples, then optimizing ground-truth samples at first frame. Specifically, devise simple tracker, composed feature extraction backbone, up-sampling decoder, multi-scale anchor-free branch. Thanks unique RMG, our FCOT can not only handle variation along temporal dimension, also overcome issue error accumulation during procedure. In addition, simplicity design, trained deployed manner real-time running speed. achieves promising performance seven benchmarks, including VOT2018, LaSOT, TrackingNet, GOT-10k, OTB100, UAV123, NFS. Code models are available at: https://github.com/MCG-NJU/FCOT.