Universal Skeleton Action Eval

---
name: universal-skeleton-action-eval
description: Evaluates a model's ability to recognize human actions from heterogeneous skeleton data with varying joint counts and topologies. It probes cross-domain generalization, zero-shot/few-shot transfer, and robustness to structural discrepancies between sensing modalities. Use when the user wants to benchmark on NTU-60, HumanML3D, NW-UCLA, NTU-120, or asks about evaluating this task. Reports Accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.17013
  bibtex_key: [REDACTED]
  confidence: high
---

# universal-skeleton-action-eval

> Towards Universal Skeleton-Based Action Recognition — Kuang et al. (2026) (arXiv:2604.17013, 2026)

## What this evaluates

Evaluates a model's ability to recognize human actions from heterogeneous skeleton data with varying joint counts and topologies. It probes cross-domain generalization, zero-shot/few-shot transfer, and robustness to structural discrepancies between sensing modalities.

## Datasets

- **NTU-60** — total ?; splits: x-sub (-1)
- **HumanML3D** — total ?; splits: Overall (-1), Many-shot (-1), Medium-shot (-1), Few-shot (-1)
- **NW-UCLA** — total ?; splits: test (-1)
- **NTU-120** — total ?; splits: x-sub (-1)

## Metrics

- `Accuracy` **(primary)** — range: percent
  - Standard classification accuracy: the proportion of correctly predicted action labels out of the total number of test samples. Evaluated across overall, many-shot, medium-shot, and few-shot categories based on action frequency, as well as subject-independent (x-sub) splits.

## Input / output format

**Input**: Skeleton sequences represented as joint coordinates, bone vectors, or motion capture data, with varying numbers of joints and topologies across datasets. Paired with text descriptions of action labels.

**Output**: Predicted action class label (or probability distribution over the open vocabulary of action classes).

## 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)) * 100
```

## Common pitfalls

- Evaluating on subject-dependent (x-see) splits instead of the standard subject-independent (x-sub) splits used in the paper.
- Ignoring the few-shot/medium-shot/many-shot partitioning for HumanML3D, which requires grouping actions by training frequency rather than evaluating overall accuracy alone.
- Assuming uniform skeleton topology; the protocol explicitly handles heterogeneous joint counts via zero-padding or interpolation, which affects input alignment.

## Evidence (verbatim from paper)

> Under a comparable protocol where the encoder is frozen and training a linear classifier, our model (w/ FC classifier) achieves 93.5% accuracy on NW-UCLA and 79.5% on NTU-120, despite not encountering NW-UCLA data during its initial training.

## Citation

```bibtex
@misc{kuang2026towards,
  title={Towards Universal Skeleton-Based Action Recognition},
  author={Kuang et al. (2026)},
  year={2026},
  note={arXiv:2604.17013}
}
```

- arXiv: 2604.17013