Hardvs2.0 Eval

---
name: hardvs2.0-eval
description: Evaluates multi-modal human activity recognition capabilities by classifying 300 action categories from synchronized RGB frames and asynchronous event streams under challenging real-world conditions such as low light, occlusion, and dynamic backgrounds. Use when the user wants to benchmark on HARDVS 2.0, or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2504.05830
  bibtex_key: wang2025hardvs
  confidence: high
---

# hardvs2.0-eval

> Human Activity Recognition using RGB-Event based Sensors: A Multi-modal Heat Conduction Model and A Benchmark Dataset — Shiao Wang et al. (arXiv:2504.05830, 2025)

## What this evaluates

Evaluates multi-modal human activity recognition capabilities by classifying 300 action categories from synchronized RGB frames and asynchronous event streams under challenging real-world conditions such as low light, occlusion, and dynamic backgrounds.

## Datasets

- **HARDVS 2.0** — total 107646; splits: train (64526), val (10734), test (32386); repo https://github.com/Event-AHU/HARDVS

## Metrics

- `accuracy` **(primary)** — range: [0, 1]
  - Standard top-1 classification accuracy: the proportion of correctly predicted class labels out of the total number of test instances.

## Input / output format

**Input**: Paired RGB video frames and asynchronous event streams captured at 346×260 resolution, with each sequence lasting approximately 5 seconds.

**Output**: A single discrete class label corresponding to one of the 300 predefined human activity categories.

## Scoring recipe

```python
correct = 0
total = len(predictions)
for pred, gold in zip(predictions, gold_labels):
    if pred == gold:
        correct += 1
return correct / total
```

## Common pitfalls

- The dataset enforces a strict 60/10/30 split per category rather than a global random split, so models must be evaluated on category-balanced test sets.
- Inputs consist of dual-modality data (RGB frames + event streams), requiring specialized fusion architectures; standard RGB-only or event-only baselines are not directly comparable.

## Evidence (verbatim from paper)

> We split 60%,10%, and 30% of each category for training, validating, and testing, respectively. In total, the number of videos in the training, validating, and testing subsets are 64526, 10734, and 32386, respectively.

## Citation

```bibtex
@misc{wang2025hardvs,
  title={Human Activity Recognition using RGB-Event based Sensors: A Multi-modal Heat Conduction Model and A Benchmark Dataset},
  author={Shiao Wang et al.},
  year={2025},
  note={arXiv:2504.05830}
}
```

- arXiv: 2504.05830