Reproduce “Unsupervised Learning of Accurate Siamese Tracking” on OpenResearch

About this paper

Qiuhong Shen, Lei Qiao, Jinyang Guo, Peixia Li, Xin Li, Bo Li, Weitao Feng, Weihao Gan and 2 more

The ULAST framework enables unsupervised training of Siamese object trackers by utilizing palindromic video sequences and novel modules to learn robust temporal correspondence. This approach achieves performance on par with some supervised methods on benchmarks like VOT2018, OTB2015, TrackingNet, and LaSOT by effectively addressing challenges related to cycle training and noisy pseudo-labels.

Key insights

• The combination of differentiable feature selection through the Region Mask and robust template propagation via CPT is critical for achieving stable and effective unsupervised cycle training in visual tracking.
• Integrating both long-term (initial template) and short-term (aggregated features from the previous frame) information within the CPT module significantly enhances template reliability and tracking robustness over time.
• Dynamically re-weighting loss functions based on the quality of pseudo-labels helps improve training performance, particularly for the box-regression branch, which is sensitive to precise object boundary information.

Summary by alphaXiv.