qhjqhj00 tnl2k-eval
ReasoningTrack: Chain-of-Thought Reasoning for Long-term Vision-Language Tracking — Xiao Wang et al. (2025) (arXiv:2508.05221, 2025)
What this evaluates
Evaluates natural language-based tracking on 2000 YouTube and surveillance videos, testing the model's ability to follow and adapt to language descriptions over time.
Datasets
- TNL2K — total 2000; splits: train (1300), test (700)
Metrics
AUC(primary) — range: percent- Area Under the Curve: area under the precision plot across varying distance thresholds.
PR— range: percent- Precision Rate: percentage of frames where the center distance is within 20 pixels.
Input / output format
Input: Video frames, initial language description, and ground-truth bounding boxes.
Output: Predicted bounding box coordinates per frame.
Scoring recipe
For each frame, compute center distance d.
PR = (count(d < 20) / N) * 100
AUC = integral of precision plot over thresholds [0, 50]
Common pitfalls
- YouTube videos often contain rapid motion and occlusions that challenge language grounding.
- Baseline methods may not be fine-tuned on TNL2K, making direct comparison sensitive to training data.
Evidence (verbatim from paper)
The dataset is divided into 1300 training videos and 700 test videos. As shown in Table[VI], our method achieves competitive performance, improving by 3% in PR over the second-ranked method.
Citation
@misc{wang2025reasoningtrack,
title={ReasoningTrack: Chain-of-Thought Reasoning for Long-term Vision-Language Tracking},
author={Xiao Wang et al. (2025)},
year={2025},
note={arXiv:2508.05221}
}
- arXiv: 2508.05221