SslTransT: Self-supervised pre-training visual object tracking with Transformers

Transformer-based visual object tracking surpasses conventional CNN-based counterparts in superior performance but comes with additional computational overhead. Existing Transformer-based trackers rely on large-scale annotated data and longer training periods. To address this issue, we introduce a self-supervised pretext task, named target localization, which randomly crops the target and then pastes it onto various background images. The copy-paste-transform data augmentation strategy can composite sufficient training data and facilitate routine training. In addition, freezing the CNN backbone during pre -training and randomly adjusting template and search area factors further lead to faster training convergence. Extensive experiments both on public tracking benchmarks and real aircraft flight test videos demonstrate that our proposed tracker SslTransT significantly outperforms the baseline performance while requiring only half the training time. Furthermore, we apply SslTransT to a 6D pose measurement system based on vision and laser ranging, achieving excellent tracking results while running in real -time.