Reproduce “General Compression Framework for Efficient Transformer Object Tracking” on OpenResearch

About this paper

Lingyi Hong, Jinglun Li, Xinyu Zhou, Shilin Yan, Pinxue Guo, Kaixun Jiang, Zhaoyu Chen, Shuyong Gao and 5 more

Researchers at Fudan University introduce CompressTracker, a general framework for compressing transformer object trackers that enhances efficiency while largely preserving accuracy. It achieves up to a 2.42x speedup with approximately 99% accuracy retention on benchmarks like LaSOT, and reduces training time to a single end-to-end step.

Key insights

• Dividing models into stages and allowing dynamic teacher-student stage replacement enables knowledge transfer to structurally diverse student architectures without complex multi-stage training.
• Progressively increasing the student's involvement during training creates an effective curriculum that fosters self-sufficiency and eliminates the need for separate fine-tuning steps.
• Combining high-level prediction guidance with fine-grained stage-wise feature mimicking ensures comprehensive knowledge transfer, leading to superior accuracy retention in compressed models.

Summary by alphaXiv.