U2flow Eval

---
name: u2flow-eval
description: Evaluates the accuracy of dense optical flow estimation and the reliability of per-pixel uncertainty quantification in an unsupervised setting. It probes the model's ability to handle occlusions, textureless regions, and domain shifts without ground-truth flow supervision. Use when the user wants to benchmark on KITTI, Sintel, or asks about evaluating this task. Reports EPE.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.10056
  bibtex_key: sun2026u2flow
  confidence: high
---

# u2flow-eval

> U$^{2}$Flow: Uncertainty-Aware Unsupervised Optical Flow Estimation — Sun et al. (2026) (arXiv:2604.10056, 2026)

## What this evaluates

Evaluates the accuracy of dense optical flow estimation and the reliability of per-pixel uncertainty quantification in an unsupervised setting. It probes the model's ability to handle occlusions, textureless regions, and domain shifts without ground-truth flow supervision.

## Datasets

- **KITTI** — total ?; splits: train (-1), test (-1)
- **Sintel** — total ?; splits: train (-1), test (-1)

## Metrics

- `EPE` **(primary)** — range: pixels
  - Average Euclidean distance between predicted and ground-truth flow vectors over all pixels.
- `Fl-all` — range: percent
  - Percentage of pixels where the endpoint error exceeds a threshold (typically 3 pixels or 5% of image diagonal).
- `AUSE` — range: [0, 1]
  - Area Under the Sparsification Error curve, measuring how well predicted uncertainty ranks true errors compared to an oracle. Lower values indicate better uncertainty calibration.
- `Spearman's CC` — range: [-1, 1]
  - Spearman's Rank Correlation Coefficient measuring the monotonic relationship between predicted uncertainty and true error. Higher values indicate stronger correlation.

## Input / output format

**Input**: Paired image frames resized to 256×832 (KITTI) or 448×1024 (Sintel), optionally with appearance and spatial augmentations.

**Output**: Dense optical flow field (H×W×2) and a per-pixel uncertainty map (H×W).

## Scoring recipe

```python
def compute_metrics(pred_flow, gt_flow, uncertainty):
    epe = np.mean(np.linalg.norm(pred_flow - gt_flow, axis=-1))
    fl_all = np.mean((np.linalg.norm(pred_flow - gt_flow, axis=-1) > 3.0) * 100)
    errors = np.linalg.norm(pred_flow - gt_flow, axis=-1)
    ause = sparsification_auc(errors, uncertainty)
    cc = spearmanr(errors, uncertainty).correlation
    return epe, fl_all, ause, cc
```

## Common pitfalls

- The paper reports Fl-all as the primary flow error metric but also breaks it down into Fl-noc and Fl-occ; readers often conflate these or miss that Fl-all is the headline number.
- AUSE is a lower-is-better metric for uncertainty quality, whereas Spearman's CC is higher-is-better; misinterpreting the directionality leads to incorrect conclusions about uncertainty reliability.
- Ablation and generalization tables report metrics on the training sets, while main benchmark comparisons use official test sets; mixing these up invalidates comparisons.

## Evidence (verbatim from paper)

> We evaluate our method using standard optical flow metrics, including the average endpoint error (EPE) and the percentage of erroneous pixels (Fl). Comparisons are conducted against both supervised and unsupervised approaches on the KITTI and Sintel benchmarks. ... We evaluate the reliability of our uncertainty estimates by measuring their correlation with the ground-truth flow error (EPE). To this end, we employ two standard quantitative metrics: the Area Under the Sparsification Error curve (AUSE) and Spearman’s Rank Correlation Coefficient (CC).

## Citation

```bibtex
@misc{sun2026u2flow,
  title={U$^{2}$Flow: Uncertainty-Aware Unsupervised Optical Flow Estimation},
  author={Sun et al. (2026)},
  year={2026},
  note={arXiv:2604.10056}
}
```

- arXiv: 2604.10056