Efficient early exit single object tracking via general distribution

Current mainstream object tracking algorithms require search region to undergo feature extraction and relation modeling through the entire backbone network. For simple scenarios with high object-background distinguishability and minimal object appearance changes, this fixed inference path design paradigm, which does not consider tracking difficulty levels, leads to redundant computations. This not only reduces inference speed but also wastes significant computational resources. To address this limitation, this work proposes an early exit object tracking algorithm based on General Distribution (GD). Leveraging the property that GD can reflect current tracking localization quality, we use the information entropy of GD as the early exit criterion, enabling differentiated processing of simple and challenging tracking scenarios. We introduce Bypass Branch Modules to enhance the prediction capability of shallow prediction heads and design a two-stage training method that maintains the performance of the final prediction head while further improving the prediction capability of shallow heads through self-distillation. Experimental results on multiple benchmark datasets demonstrate that our proposed algorithm not only achieves accuracy metrics comparable to current state-of-the-art methods but also significantly improves inference speed. Through quantitative analysis on the LaSOT dataset test set, predictions obtained from early exits at shallow heads exhibit good prediction accuracy, which not only demonstrates that our method can effectively distinguish between detection and challenging scenes, but also validates the rationality of applying early inference termination for simple scenes.