End-to-end tracking without the post-processing hangover

What it does

STARK is the official PyTorch implementation of an ICCV 2021 visual object tracker. It directly predicts a single bounding box per frame and deliberately avoids hyperparameter-sensitive post-processing, which the authors credit for stable performance. The codebase is built on the PyTracking and DETR ecosystems and targets standard tracking benchmarks.

The interesting bit

The project treats tracking as a direct prediction task: the model outputs one bounding box and forgoes hyperparameter-sensitive post-processing entirely. A lightweight offshoot called STARK-Lightning is advertised at 200–300 FPS on an RTX TITAN, beating DiMP50 while using a smaller model than SiamFC—though the README is thin on how those gains are realized.

Key highlights

• End-to-end pipeline: one bounding box out, zero hyperparameter-tuned post-processing.
• Real-time inference: 40 FPS for STARK-ST50 and 30 FPS for STARK-ST101 on a Tesla V100.
• Strong benchmark scores: 67.1 AUC on LaSOT, 68.8 AO on GOT-10K, and 82.0 AUC on TrackingNet, ahead of TransT, TrDiMP, and Siam R-CNN.
• Competition results: winner of the VOT-21 RGB-D challenge and runner-up in the real-time and long-term tracks.
• Upstream adoption: integrated into the OpenMMLab mmtracking library.

Caveats

• The README is essentially a quickstart manual; architectural details and training intuition are left to the paper PDF.
• The STARK-Lightning speed claims are mentioned without detail on the speedup mechanism or reproducibility notes in the main README.

Verdict

A solid choice for researchers who want a real-time, PyTorch-native tracking baseline with strong benchmark numbers. Skip it if you need multi-object tracking or extensive documentation beyond training recipes.