Transformers for Single Object Tracking: Temporal Context Propagation and Frame Relationship Modeling

This paper primarily proposes an optimization scheme for single object tracking based on the Transformer model. The paper puts forward a model optimization scheme based on the Siamese, DiMP, and TrDiMP baselines. In the tracking process, we accurately predict the target center by densely calculating confidence scores in a wider search area. By integrating various target representations and spatial clues, we introduce a Local Perception Unit (LPU) to extract local information, simplifying the training process of object search. In the Transformer decoder of our proposed model, which is responsible for managing search features and previously stored template content, the decoder effectively transmits key temporal information between frames. By propagating features from previous frames to the current frame, we effectively mitigate the impact of appearance changes and background noise. The focus is on validating the efficacy of lightweight multi-head self-attention in the model, while also utilizing spatial information to reduce parameters and enhance the model’s efficiency.