Appearance Free Action Eval

---
name: appearance-free-action-eval
description: Evaluates zero-shot generalization of action recognition models to appearance-free videos generated by warping noise or random dots with optical flow, testing reliance on motion cues over static shape and texture. Use when the user wants to benchmark on UCF5, AFD5, AFF5, or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.16675
  bibtex_key: kumar2026appearancefree
  confidence: high
---

# appearance-free-action-eval

> Appearance-free Action Recognition: Zero-shot Generalization in Humans and a Two-Pathway Model — Prerana Kumar, Martin A. Giese (2026) (arXiv:2604.16675, 2026)

## What this evaluates

Evaluates zero-shot generalization of action recognition models to appearance-free videos generated by warping noise or random dots with optical flow, testing reliance on motion cues over static shape and texture.

## Datasets

- **UCF5** — total ?; splits: train (-1), test (-1)
- **AFD5** — total ?; splits: test (-1)
- **AFF5** — total ?; splits: test (-1)

## Metrics

- `accuracy` **(primary)** — range: [0, 1]
  - Fraction of correctly classified video instances out of the total number of instances in a given split.
- `Transfer Score` — range: [0, 1]
  - Mean of accuracy on AFD5 and AFF5: (AFD5_acc + AFF5_acc) / 2.

## Input / output format

**Input**: RGB video frames or optical flow sequences representing action clips.

**Output**: Discrete action class label.

## Scoring recipe

```python
def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return correct / len(gold_labels)

def compute_transfer_score(acc_afd5, acc_aff5):
    return (acc_afd5 + acc_aff5) / 2
```

## Common pitfalls

- Models with explicit optical flow inputs still suffer large performance drops on appearance-free videos without a coherence-gating mechanism.
- RGB-only models collapse to near-chance performance on appearance-free stimuli, highlighting the necessity of motion cues.
- Human performance is reported as a baseline but is not used to train or tune the models.

## Evidence (verbatim from paper)

> Mean accuracy across participants was $0.9891\pm 0.0146$ on UCF5, $0.8400\pm 0.0613$ on AFD5, and $0.7872\pm 0.0539$ on AFF5. Transfer Score is $(\mathrm{AFD5}+\mathrm{AFF5})/2$ computed from the reported means .

## Citation

```bibtex
@misc{kumar2026appearancefree,
  title={Appearance-free Action Recognition: Zero-shot Generalization in Humans and a Two-Pathway Model},
  author={Prerana Kumar, Martin A. Giese (2026)},
  year={2026},
  note={arXiv:2604.16675}
}
```

- arXiv: 2604.16675