Lightweight Attention-Enhanced Adapters for Visual Drone Tracking

In this paper, we present a novel fine-tuning method for visual drone tracking named AdtTrack. The method achieves efficient fine-tuning on drone tracking datasets by introducing adapter modules into the base tracking model. Unlike traditional full-parameter fine-tuning methods that rely on large-scale general datasets and high-performance GPUs, AdtTrack adopts a simple adapter fine-tuning approach. It introduces only a small number of trainable parameters in multi-layer adapters with the remaining parameters frozen and adds an attention mechanism to the adapter module of one encoder layer to enhance key feature representation capabilities. This approach not only reduces dependency on large-scale training data but also significantly decreases training time and memory consumption. Experiments demonstrate that under the same training conditions, the AdtTrack-256 model improves the area under the receiver operating characteristic curve (AUC) by 5.3% compared with full-parameter fine-tuned models on drone tracking datasets, effectively avoiding the problem of catastrophic forgetting.