Kandinsky LabKandinsky video analysis
Kandinsky Video Motion Predictor
This trained analysis model adapts VideoMAEv2-Base. It predicts three motion-intensity scores to help filter and review video datasets.
Research release
A regression head learns from human labels
Training used 5,000 annotated clips and updated both the backbone and regression head. The three output scores cover camera movement, object movement and other dynamics, each on a zero-to-three scale.Kandinsky video tools: trained motion analysis models
Camera and object motion can be confused
The developers note ambiguity between viewpoint movement and object movement. The other-dynamics score also responds to scene transitions, which can help identify slideshows.Kandinsky video tools: trained motion analysis models
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | ai-forever/kandinsky-video-motion-predictor |
| Base model | OpenGVLab/VideoMAEv2-Baseai-forever/kandinsky-video-motion-predictor — model card and files |
| Weight license | Not stated in the checked model card.ai-forever/kandinsky-video-motion-predictor — model card and files |