qhjqhj00 chiral-action-recognition-eval
Chirality in Action: Time-Aware Video Representation Learning by Latent Straightening — Bagad et al. (2025) (arXiv:2509.08502, 2025)
What this evaluates
Evaluates a video representation's sensitivity to temporal direction by distinguishing between temporally opposite actions (e.g., opening vs. closing a door). Also tests general action recognition capability via linear probing on standard benchmarks.
Datasets
- Something-Something v2 — total ?; splits: test (-1)
- EPIC-Kitchens — total ?; splits: test (-1)
- Charades — total ?; splits: test (-1)
- Kinetics-400 — total ?; splits: test (-1)
- UCF-101 — total ?; splits: test (-1)
- HMDB-51 — total ?; splits: test (-1)
Metrics
Chiral Accuracy(primary) — range: percent- Percentage of correctly classified chiral action pairs by a linear classifier trained on video descriptors.
Input / output format
Input: Video clips sampled at T=16 frames linearly. Per-frame features extracted from an image encoder and concatenated or processed into a single video descriptor.
Output: Predicted class label for the chiral action pair or standard action class.
Scoring recipe
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
- Naively concatenating per-frame features yields bulky descriptors that may be impractical for large-scale indexing, despite competitive performance.
- Average pooling over space and time can lose temporal sensitivity compared to average pooling over space and concatenating over time.
- Transformer-based models with a CLS token should be handled carefully (use CLS or average pool tokens), as treating them differently affects descriptor dimensionality.
Evidence (verbatim from paper)
From the results shown in Table 2, we observe that the proposed LiFT features achieve the best performance on SSv2, EPIC and Charade, while being compact (d=768). ... Chiral Accuracy ↑
Citation
@misc{bagad2025chirality,
title={Chirality in Action: Time-Aware Video Representation Learning by Latent Straightening},
author={Bagad et al. (2025)},
year={2025},
note={arXiv:2509.08502}
}
- arXiv: 2509.08502