Smaller and more Accurate Swin-transformer Model Prediction for Tracking
Fei Pan, Lianyu Zhao, Chenglin Wang · 2023
Recently, Transformers have been excelling in the field of object tracking. However, most existing trackers tend to indiscriminately increase the number of model parameters in order to enhance tracking accuracy. In contrast, we introduce the Swin-Transformer-Tiny-based Predictive Model for Tracking (STMP). The STMP leverages Transformers for feature extraction and fusion, combines positional encoding and neighboring candidate information, and redesigns the model predictor to accurately locate and estimate target bounding boxes. This renders STMP as a compact yet powerful tracker, with only 32.56M model parameters. Evaluation experiments show that our STMP performs outstandingly on five challenging datasets. Particularly in the GOT-10k evaluation, our Average Overlap (AO) scored an excellent 75.2%, not only far surpassing trackers with similar levels of model parameters but also exceeding most larger model trackers. Simultaneously, we achieved commendable scores of 82.3% and 68.8% on the TrackingNet and LaSOT datasets respectively.