DCFNet++: More Advanced Correlation Filters Network for Real-Time Object Tracking

Visual object tracking has been widely addressed in Siamese networks, where accurate and fast object tracking can be achieved. However, it is challenging to discriminate foregrounds from semantic backgrounds, because the semantic backgrounds are always considered as distractors, which would hinder the robustness of Siamese trackers. In this paper, we address the visual object tracking problem in complex scenarios, including occlusions, out-of-view, deformation, background cluttering, and other variations. We introduce multiple combinations in the training set to improve the discriminative ability of the learned features. As a result, the robustness of the model can be improved. We also propose an advanced method to fuse multi-layer features, so that the feature representation can be enhanced. Finally, we develop a flow-based tracking method, which makes the tracker very robust to occlusion scenarios. Extensive evaluations on OTB-2015, VOT2018, UAV123, and LaSOT benchmarks demonstrate that the proposed DCFNet++ has strong robustness when facing challenging scenarios. Without bells and whistles, our proposed tracker can run at more than 56 FPS during test time.