Adjust Sample Imbalance and Exclude Similar Object in Underwater Object Tracking

Although modern trackers have competitive performance when dealing with underwater image degradation, there are still two problems when applying them to Underwater Object Tracking (UOT). On the one hand, the single object tracker is trained on the open-air datasets, which means that the tracker has a serious sample imbalance between underwater objects and open-air objects when applied to UOT. On the other hand, underwater targets such as fish and dolphins usually have a similar appearance, it is challenging for the model itself to discriminate the weak discriminative features. The existing detection-based post processing is hard to distinguish the tracked target among similar objects. In this paper, we propose UOSTrack, which consists of Underwater images and Open-air sequences Hybrid Training (UOHT) and Motion-based Post Processing (MBPP). UOHT is designed to adjust the sample imbalance underwater tracker. Specifically, Underwater Object Detection (UOD) image is converted into imag pairs through customized data augmentation, so that the tracker has more underwater domain training samples and learn the feature expression of underwater objects. MBPP is proposed to exclude similar objects around the target. Specifically, it uses the estimation box predicted by the Kalman Filter and candidate boxes in each frame to reconfirm the target that is hidden in the candidate area when the target is lost. UOSTrack has an average performance improvement of 3.5% over OSTrack on Similar Object challenge of the UOT100 and UTB180 datasets. The average performance improvement of UOSTrack on UOT100 and UTB180 is 1% and 3%, respectively. Experiments on two UOT benchmarks demonstrate the effectiveness of UOHT and MBPP, and the generalization and applicability of MBPP for UOT.