Ntu Rgb D Eval

ntu-rgb-d-eval

Skeleton Focused Human Activity Recognition in RGB Video — Yu et al. (2020) (arXiv:2004.13979, 2020)

What this evaluates

Evaluates a model's ability to recognize human activities from RGB videos by leveraging skeleton-driven attention to focus on spatial-temporal regions of interest. It measures classification accuracy under standard cross-subject and cross-view protocols.

Datasets

• NTU-RGB+D — total ?; splits: cross-subject (-1), cross-view (-1)
• Northwestern-UCLA Multiview — total ?; splits: cross-view (-1)

Metrics

• accuracy (primary) — range: percent
  • Percentage of correctly classified video samples out of the total number of samples in the evaluation split.

Input / output format

Input: RGB video frames and corresponding skeleton joint coordinates for each frame.

Output: Predicted activity class label.

Scoring recipe

correct = sum(1 for pred, gold in zip(predictions, gold_labels) if pred == gold)
accuracy = (correct / len(gold_labels)) * 100
return accuracy

Common pitfalls

• Must strictly follow the official cross-subject and cross-view train/test splits for NTU-RGB+D.
• Ensemble methods require freezing the GCN skeleton encoder weights during RGB branch training and evaluation to match reported results.
• Accuracy is reported as a percentage, not a decimal fraction.

Evidence (verbatim from paper)

In Fig. 6, we illustrated the effectiveness of the weighted ST-ROI method that improves the recognition accuracy of every activity of Northwestern-UCLA Dataset.

Citation

@misc{yu2020skeleton,
  title={Skeleton Focused Human Activity Recognition in RGB Video},
  author={Yu et al. (2020)},
  year={2020},
  note={arXiv:2004.13979}
}

• arXiv: 2004.13979