Realtime object tracking models

The current state-of-the-art tracking models are hampered by low speed, limiting their applicability on devices with limited computational power. Although Existing realtime object tracking models could reach high speeds on edge devices, their performance is poor. Consequently, the high-performance tracker with fast speed on edge devices is critical. I have two plans to solve this problem. The first Scheme is to use the pre-trained lightweight transformer as the tracking models' backbone and through the proper design, the tracking model not only has good performance because of the use of the pre-trained visual transformer but also because the visual transformer used is lightweight and does not slow down the model due to the huge amount of computation caused by the transformer. The second Scheme is to build a small tracking model and pre-train the small tracking model using the MAE pre-training method, and finally, choose the current best-performing transformer-based tracking model as the teacher model to train the student model. Eventually, two state-of-the-art real-time tracking models will be provided, along with their code, and they will be maintained over time.