Reproduce “LiteTrack: Layer Pruning with Asynchronous Feature Extraction for Lightweight and Efficient Visual Tracking” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Qingmao Wei, Bi Zeng, Jianqi Liu, Li He, Guotian Zeng

LiteTrack introduces a lightweight and efficient visual tracker that combines systematic layer pruning with asynchronous feature extraction for transformer models. This approach enables real-time operation on edge devices, achieving 64 FPS on a Jetson Orin NX while maintaining competitive accuracy with non-real-time state-of-the-art trackers.

Key insights

  • The synergistic combination of systematic layer pruning and asynchronous feature extraction can not only counteract performance degradation but can also enhance accuracy compared to heavier, unoptimized baselines.
  • Transformer-based tracking models can efficiently integrate template feature caching, traditionally an advantage of two-stage architectures, while retaining the powerful end-to-end learning capabilities of one-stage methods.
  • Scaling down powerful, pre-trained heavy models through intelligent architectural modifications provides a more effective route to achieving high-accuracy, real-time tracking on edge devices than designing lightweight models from scratch.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.