README.md

SimMotion-Real Benchmark (SemanticMoments)

This folder contains the real-world benchmark for evaluating motion representation consistency as described in:

"SemanticMoments: Training-Free Motion Similarity via Third Moment Features"

Dataset Structure

The benchmark consists of 40 real-world test cases, each organized as a triplet:

1. ref.mp4: The reference video defining the target semantic motion.
2. positive.mp4: A real-world video sharing the same semantic motion as the reference.
3. negative.mp4: Hard Negative. A video with a similar visual appearance to the reference, but containing a different motion.

Evaluation Protocol

This benchmark evaluates retrieval performance using a large-scale distractor pool to ensure motion representations are robust against diverse real-world actions.

• Retrieval Pool: For each reference video, the candidate pool includes:
  • The corresponding positive.mp4.
  • The corresponding hard negative.mp4.
  • 1,000 Distractor Videos: The first 1,000 videos from the Kinetics-400 validation set (sorted alphabetically).
• Metric: Top-1 Accuracy. A trial is successful if: $$Similarity(ref,positive)>Similarity(ref,candidate{s}_{all})Similarity(ref,\; positive)\; >\; Similarity(ref,\; candidates\_\{all\})$$ (Where $candidates_{all}$ includes the hard negative and all 1,000 Kinetics distractors).

Citation

If you use this benchmark, please cite:

@article{huberman2026semanticmoments,
  title={SemanticMoments: Training-Free Motion Similarity via Third Moment Features},
  author={Huberman, Saar and Goldberg, Kfir and Patashnik, Or and Benaim, Sagie and Mokady, Ron},
  journal={arXiv preprint arXiv:2602.09146},
  year={2026}
}