Omniflow Eval

---
name: omniflow-eval
description: Evaluates optical flow estimation models on synthetic omnidirectional human motion data. It probes the model's ability to handle fisheye distortions, domain-randomized environments, and varying amounts of fine-tuning data. Use when the user wants to benchmark on OmniFlow, or asks about evaluating this task. Reports optical flow error.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2104.07960
  bibtex_key: seidel2021omniflow
  confidence: medium
---

# omniflow-eval

> OmniFlow: Human Omnidirectional Optical Flow — Seidel et al. (2021) (arXiv:2104.07960, 2021)

## What this evaluates

Evaluates optical flow estimation models on synthetic omnidirectional human motion data. It probes the model's ability to handle fisheye distortions, domain-randomized environments, and varying amounts of fine-tuning data.

## Datasets

- **OmniFlow** — total ?; splits: test (-1)

## Metrics

- `optical flow error` **(primary)** — range: other
  - Not explicitly stated in the provided section; optical flow benchmarks typically report End-Point Error (EPE) or F1 score.

## Input / output format

**Input**: Pairs of omnidirectional (fisheye) images $I_1, I_2$.

**Output**: Per-pixel optical flow field representing displacement between $I_1$ and $I_2$.

## Scoring recipe

```python
Not explicitly provided in the text. Standard optical flow evaluation computes the mean Euclidean distance between predicted and ground-truth flow vectors:
```python
def compute_metric(pred, gold):
    error = np.sqrt((pred[:,:,0] - gold[:,:,0])**2 + (pred[:,:,1] - gold[:,:,1])**2)
    return np.mean(error)
```
```

## Common pitfalls

- Test-time augmentation (TTA) is applied during testing due to the lack of alternative omnidirectional flow datasets, which may inflate reported performance.
- Model performance is highly sensitive to fine-tuning data size; 5k pairs suffice for RAFT, but correlation-based CNNs require ~20k pairs.
- Input resolution is resized to 512x512 during training, potentially affecting high-resolution flow accuracy.

## Evidence (verbatim from paper)

> Our dataset is evaluated on a test set of OmniFlow of 10% of the whole dataset. We train a correspondence network for optical flow and fine-tune on five subsets 1k, 5k, 10k, 15k and 20k of OmniFlow with a pretrained model on FlyingChairs and FlyingThings. As long there is no further omnidirectional optical flow dataset for testing available we use test-time augmentation (TTA) with three standard augmentation methods cropping, scaling and horizonal flipping. Results on OmniFlow test set are shown in Figure 2.

## Citation

```bibtex
@misc{seidel2021omniflow,
  title={OmniFlow: Human Omnidirectional Optical Flow},
  author={Seidel et al. (2021)},
  year={2021},
  note={arXiv:2104.07960}
}
```

- arXiv: 2104.07960