Few Shot Action Recognition Eval

---
name: few-shot-action-recognition-eval
description: Evaluates a model's ability to recognize human actions in video clips using only a few labeled examples per class. It probes the model's capacity to leverage motion dynamics and semantic cues for robust classification under data-scarce conditions. Use when the user wants to benchmark on Something-Something, Kinetics, UCF101, HMDB51, FineGym, or asks about evaluating this task. Reports average few-shot accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2508.03695
  bibtex_key: [REDACTED]
  confidence: high
---

# few-shot-action-recognition-eval

> Trokens: Semantic-Aware Relational Trajectory Tokens for Few-Shot Action Recognition — Pulkit Kumar et al. (2025) (arXiv:2508.03695, 2025)

## What this evaluates

Evaluates a model's ability to recognize human actions in video clips using only a few labeled examples per class. It probes the model's capacity to leverage motion dynamics and semantic cues for robust classification under data-scarce conditions.

## Datasets

- **Something-Something** — total ?; splits: Small (10000), Full (-1)
- **Kinetics** — total ?; splits: few-shot (-1)
- **UCF101** — total ?; splits: few-shot (-1)
- **HMDB51** — total ?; splits: few-shot (-1)
- **FineGym** — total ?; splits: few-shot (-1)

## Metrics

- `average few-shot accuracy` **(primary)** — range: percent
  - Accuracy averaged over 10,000 evaluation episodes. Each episode follows a standard N-way K-shot protocol where the model classifies query videos into one of N classes using K support examples per class.

## Input / output format

**Input**: Video clips representing human actions, processed through a DINOv2 feature extractor and CoTracker to generate semantic-aware trajectory points.

**Output**: Predicted class label for each query video clip.

## Scoring recipe

```python
accuracy = 0
for episode in 10000_episodes:
  support_set = sample(K examples per class from N classes)
  query_set = sample remaining examples
  predictions = model.predict(query_set, support_set)
  correct = sum(predictions == ground_truth)
  accuracy += correct / len(query_set)
return accuracy / 10000 * 100
```

## Common pitfalls

- Kinetics dataset exhibits strong appearance bias, which can mask the effectiveness of motion-focused modules if not accounted for.
- FLOP efficiency comparisons exclude the DINO feature extractor, so reported computational costs are relative and not absolute.
- Evaluation strictly follows few-shot splits from prior literature; deviating from these splits breaks comparability with reported baselines.

## Evidence (verbatim from paper)

> Following standard protocols, we evaluate using average few-shot accuracy across 10,000 episodes. We evaluate Trokens against previous state-of-the-art approaches under the standard 5-way K-shot setting.

## Citation

```bibtex
@misc{kumar2025trokens,
  title={Trokens: Semantic-Aware Relational Trajectory Tokens for Few-Shot Action Recognition},
  author={Pulkit Kumar et al. (2025)},
  year={2025},
  note={arXiv:2508.03695}
}
```

- arXiv: 2508.03695