Tnllt Eval

---
name: tnllt-eval
description: Evaluates long-term vision-language tracking capability by measuring localization accuracy over extended video sequences while dynamically updating natural language descriptions to handle appearance changes and occlusions. Use when the user wants to benchmark on TNLLT, or asks about evaluating this task. Reports PR.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2508.05221
  bibtex_key: [REDACTED]
  confidence: high
---

# tnllt-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 long-term vision-language tracking capability by measuring localization accuracy over extended video sequences while dynamically updating natural language descriptions to handle appearance changes and occlusions.

## Datasets

- **TNLLT** — total 200; splits: test (-1)

## Metrics

- `PR` **(primary)** — range: percent
  - Precision Rate: percentage of frames where the center distance between predicted and ground-truth boxes is within 20 pixels.
- `NPR` — range: percent
  - Normalized Precision Rate: percentage of frames where the center distance is within a dynamic threshold scaled to the ground-truth object size.
- `SR` — range: percent
  - Success Rate: percentage of frames where the Intersection-over-Union (IoU) between predicted and ground-truth boxes exceeds 0.5.

## Input / output format

**Input**: Video frames (search and template), initial language description, and ground-truth bounding boxes.

**Output**: Updated language description and predicted bounding box coordinates per frame.

## Scoring recipe

```python
For each frame, compute center distance d between predicted and GT box.
PR = (count(d < 20) / N) * 100
NPR = (count(d < threshold) / N) * 100
SR = (count(IoU(pred, GT) > 0.5) / N) * 100
```

## Common pitfalls

- Text update interval significantly impacts accuracy; optimal interval is dataset-dependent (e.g., 100 frames on TNLLT).
- Cascaded text updates can suffer from error accumulation, making initial text sometimes more robust than refined text.
- Static text integration can improve robustness compared to purely dynamic updates.

## Evidence (verbatim from paper)

> The dataset is evaluated using three metrics: precision (P), normalize precision(NP) and success rate(SR) metrics. As shown in Table[III], our method has achieved the best results among the various approaches in recent years, with an PR of 74.1%, NPR of 77.0%, SR of 63.9%.

## Citation

```bibtex
@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