SemanticMoments: Training-Free Motion Similarity via Third Moment Features
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866917262813822976 |
|---|---|
| author | Huberman, Saar Goldberg, Kfir Patashnik, Or Benaim, Sagie Mokady, Ron |
| author_facet | Huberman, Saar Goldberg, Kfir Patashnik, Or Benaim, Sagie Mokady, Ron |
| contents | Retrieving videos based on semantic motion is a fundamental, yet unsolved, problem. Existing video representation approaches overly rely on static appearance and scene context rather than motion dynamics, a bias inherited from their training data and objectives. Conversely, traditional motion-centric inputs like optical flow lack the semantic grounding needed to understand high-level motion. To demonstrate this inherent bias, we introduce the SimMotion benchmarks, combining controlled synthetic data with a new human-annotated real-world dataset. We show that existing models perform poorly on these benchmarks, often failing to disentangle motion from appearance. To address this gap, we propose SemanticMoments, a simple, training-free method that computes temporal statistics (specifically, higher-order moments) over features from pre-trained semantic models. Across our benchmarks, SemanticMoments consistently outperforms existing RGB, flow, and text-supervised methods. This demonstrates that temporal statistics in a semantic feature space provide a scalable and perceptually grounded foundation for motion-centric video understanding. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_09146 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | SemanticMoments: Training-Free Motion Similarity via Third Moment Features Huberman, Saar Goldberg, Kfir Patashnik, Or Benaim, Sagie Mokady, Ron Computer Vision and Pattern Recognition Retrieving videos based on semantic motion is a fundamental, yet unsolved, problem. Existing video representation approaches overly rely on static appearance and scene context rather than motion dynamics, a bias inherited from their training data and objectives. Conversely, traditional motion-centric inputs like optical flow lack the semantic grounding needed to understand high-level motion. To demonstrate this inherent bias, we introduce the SimMotion benchmarks, combining controlled synthetic data with a new human-annotated real-world dataset. We show that existing models perform poorly on these benchmarks, often failing to disentangle motion from appearance. To address this gap, we propose SemanticMoments, a simple, training-free method that computes temporal statistics (specifically, higher-order moments) over features from pre-trained semantic models. Across our benchmarks, SemanticMoments consistently outperforms existing RGB, flow, and text-supervised methods. This demonstrates that temporal statistics in a semantic feature space provide a scalable and perceptually grounded foundation for motion-centric video understanding. |
| title | SemanticMoments: Training-Free Motion Similarity via Third Moment Features |
| topic | Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2602.09146 |