Kinect Action Recognition Eval

kinect-action-recognition-eval

Analysis and Evaluation of Kinect-based Action Recognition Algorithms — Lei Wang (2021) (arXiv:2112.08626, 2021)

What this evaluates

Evaluates the robustness of Kinect-based action recognition algorithms across single-view and cross-view scenarios. Probes how well models handle viewpoint variation, motion variability, and different sensor modalities (depth, skeleton, RGB-D) on standardized benchmarks.

Datasets

• MSRAction3D Dataset — total ?; splits: train (-1), test (-1)
• 3D Action Pairs Dataset — total ?; splits: train (-1), test (-1)
• Cornell Activity Dataset (CAD-60) — total ?; splits: train (-1), test (-1)
• UWA3D Single View Dataset — total ?; splits: train (-1), test (-1)
• UWA3D Multiview Dataset — total ?; splits: train (-1), test (-1)

Metrics

• average recognition accuracy (primary) — range: [0, 1]
  • Calculated per class as the number of correctly predicted labels for that class divided by the total number of ground-truth labels for that class in the dataset. Overall performance is typically summarized via a confusion matrix or averaged across classes.

Input / output format

Input: Depth video sequences, 3D skeleton joint coordinates, and/or RGB-D frames captured by Kinect v1.

Output: Predicted action class label for each input sequence or frame.

Scoring recipe

def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    total = len(gold_labels)
    return correct / total if total > 0 else 0.0

Common pitfalls

• Spatiotemporal cell sizes and desired frame counts for skeleton representation vary significantly across datasets and must be manually adjusted per dataset to match original implementations.
• The HDG algorithm requires combining four specific feature vectors (hod, hodg, jpd, jmv) rather than using them individually, and its original code was unavailable, requiring custom implementation.
• Cross-view recognition strategies are only applicable to the UWA3D Multiview Dataset, while the other four datasets are strictly single-view.

Evidence (verbatim from paper)

Average recognition accuracy for a given class A is given by average recognition accuracy = (# correct class A labels from algorithm) / (# total class A labels in dataset). The measure given in Equation 4.1 is for one class only. To illustrate the recognition accuracy of an algorithm, a confusion matrix shown the recognition accuracy of all classes is often used.

Citation

@misc{wang2021kinectactionrecognition,
  title={Analysis and Evaluation of Kinect-based Action Recognition Algorithms},
  author={Lei Wang (2021)},
  year={2021},
  note={arXiv:2112.08626}
}

• arXiv: 2112.08626