# Ueof Eval

> Evaluates the accuracy of event-based optical flow estimation models in underwater environments. It probes how well algorithms handle low-texture, turbid, and refractive conditions compared to terrestrial benchmarks. Use when the user wants to benchmark on UEOF, or asks about evaluating this task. Reports AEE.

- Skill: `qhjqhj00/ueof-eval` (Agent Skill)
- Install (CLI): `npx skillmds add qhjqhj00/ueof-eval`
- Raw SKILL.md: https://api.skillmd.com/api/skills/qhjqhj00/ueof-eval/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: qhjqhj00 (https://skillmd.com/u/qhjqhj00)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/qhjqhj00/ueof-eval

---


# ueof-eval

> UEOF: A Benchmark Dataset for Underwater Event-Based Optical Flow — Truong et al. (2026) (arXiv:2601.10054, 2026)

## What this evaluates

Evaluates the accuracy of event-based optical flow estimation models in underwater environments. It probes how well algorithms handle low-texture, turbid, and refractive conditions compared to terrestrial benchmarks.

## Datasets

- **UEOF** — total ?; splits: test (5)

## Metrics

- `AEE` **(primary)** — range: pixels
  - Average Endpoint Error in pixels. Computed as the mean Euclidean distance between predicted and ground-truth optical flow vectors across all pixels.
- `ANPE@N` — range: percent
  - N-pixel error rate. Represents the percentage of pixels with an endpoint error lower than N pixels for N ∈ {1, 2, 3, 5, 10, 20}. Lower values indicate better performance.

## Input / output format

**Input**: Event stream data (and optionally RGB frames or camera ego-velocity for multimodal baselines). Models process accumulated event windows over time.

**Output**: 2D optical flow vectors (u, v) per pixel for each frame in the sequence.

## Scoring recipe

```python
def compute_metrics(pred_flow, gt_flow):
    epe = np.linalg.norm(pred_flow - gt_flow, axis=-1)
    aee = np.mean(epe)
    anpe = {}
    for n in [1, 2, 3, 5, 10, 20]:
        anpe[f'ANPE@{n}'] = np.mean(epe > n) * 100
    return aee, anpe
```

## Common pitfalls

- Models pretrained on terrestrial datasets (e.g., DSEC) suffer severe feature distribution shifts in underwater scenes due to low-frequency textures and turbidity.
- Contrast maximization methods fail under refractive caustics, mistaking light patterns for motion events.
- Deep-water scenes introduce haloing, attenuation, and low contrast, significantly increasing error rates compared to shallow-water scenes.

## Evidence (verbatim from paper)

> The event-based optical flow metrics include the average endpoint error (AEE) as well as ANPE, which represents the percentage of pixels with an endpoint error lower than N pixels for N ∈ {1, 2, 3, 5, 10, 20}.

## Citation

```bibtex
@misc{truong2026ueof,
  title={UEOF: A Benchmark Dataset for Underwater Event-Based Optical Flow},
  author={Truong et al. (2026)},
  year={2026},
  note={arXiv:2601.10054}
}
```

- arXiv: 2601.10054