Towards Grand Unification of Object Tracking

We present a unified method, termed Unicorn, that can simultaneously solve four tracking problems (SOT, MOT, VOS, MOTS) with single network using the same model parameters. Due to fragmented definitions of object problem itself, most existing trackers are developed address or part tasks and over-specialize on characteristics specific tasks. By contrast, Unicorn provides solution, adopting input, backbone, embedding, head across all For first time, we accomplish great unification architecture learning paradigm. performs on-par better than its task-specific counterparts in 8 datasets, including LaSOT, TrackingNet, MOT17, BDD100K, DAVIS16-17, MOTS20, BDD100K MOTS. believe will serve as solid step towards general vision model. Code is available at https://github.com/MasterBin-IIAU/Unicorn .