IMSiam: IoU-aware Matching-adaptive Siamese network for object tracking

Fully convolutional Siamese networks have shown their advantages in the visual object tracking task. However, most existing Siamese-based trackers still suffer from poor matching information and a low correlation between the classification score and bounding box estimation. To address these issues, we propose an IoU-aware Matching-adaptive Siamese network (IMSiam) for visual tracking in this paper. Specifically, a matching-adaptive network is proposed, which integrates multiple types of encoded feature maps and adaptively sample matching information to simultaneously perform target classification and bounding box regression. And we introduce an IoU-aware head to equip the Siamese detector with the capacity of IoU prediction for each regressed box, which then penalizes the classification score to obtain a more confident target region. Benefiting from the adaptive matching information and IoU-based classification outputs, the model can localize the target more accurately while avoiding the redundant search of scale penalty factor during tracking. Experimental results on OTB2013, OTB2015, VOT2018, VOT2019, LaSOT, and Got10k datasets demonstrate that the proposed tracker performs favorably against state-of-the-art methods.