A flow-guided self-calibration Siamese network for visual tracking

Existing Siamese-based trackers pay attention to applying online updates to deal with target deformation and occlusion. Despite excellent accuracy and robustness, these trackers are still plagued by model drift due to the cumulative errors from tracking results. Therefore, this paper proposes a flow-guided self-calibration Siamese network to alleviate the model drift. This network focuses on leveraging the rich motion information in the current frame and adaptively optimizing the feature representation of the target. Firstly, to mitigate the lack of motion information during tracking, the optical flow field is estimated before the target location. A gather network is designed carefully to extract the deep motion features from the optical flow. Then, owing to the cumulative errors caused by the tracking result, a novel self-calibration module is introduced to update the appearance model without any tracking results adaptively. The module incorporates appearance features and deep motion features via an attention mechanism. Finally, a synthetic loss function is proposed to obtain expressive deep feature representation by adding a competitive loss between samples to the original loss function. Extensive experiments have demonstrated the effectiveness of the proposed method on benchmarks.