SkipTrack: Learning Essential Skip Connections for Robust Single Object Tracking

Single object tracking plays a crucial role in numerous applications across critical fields, yet it remains one of the most demanding tasks in computer vision. The information flow in the model forms a vital part of the architecture. In the context of one-stream transformer single object tracking, information flow has been sequential only. In this work, we introduce high-level, learnable skip connections that provide skip paths over the transformer blocks. The model learns to weigh the embeddings from the sequential and skip paths to combine them into one tensor, resulting in a simpler learning objective. We demonstrate the effectiveness of our method on the leading single object tracking dataset of GOT-10K. Our model achieves state-of-the-art performance on the mentioned dataset.