Research map / 3D vision / Depth and stereo
EM-Driven Unsupervised Learning for Efficient Motion Segmentation
Etienne Meunier, Anaïs Badoual, Patrick Bouthémy
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use quadratic motion model instead of affine (ablation) | EM-driven network · mean J over four VOS datasets | Jaccard J: 53.1 -> 61.1 | Improved |
| add optical-flow data augmentation with random parametric global motion (ablation) | EM-driven network · mean J over four VOS datasets | Jaccard J: 58.4 -> 61.1 | Improved |
| use L1 loss instead of squared L2 (ablation) | EM-driven network · mean J over four VOS datasets | Jaccard J: 58.8 -> 61.1 | Improved |
| use L1 loss instead of L2 norm (ablation) | EM-driven network · mean J over four VOS datasets | Jaccard J: 60.6 -> 61.1 | Improved |
| increase number of masks K from 3 to 6 | EM-driven network · DAVIS2016 val | Jaccard J: 75.1 -> 78.3 | Improved |
| increase number of masks K from 6 to 7 | EM-driven network · DAVIS2016 val | Jaccard J: 78.3 -> 78.3 | No clear effect |
| use trained network instead of classical EM with one random initialisation | EM segmentation · DAVIS2016 val | Jaccard J: 59.0 -> 69.3 | Improved |
| use trained network instead of classical EM with best of 500 initialisations | EM segmentation · DAVIS2016 val | inference time per frame (ms): 4800 -> 8 | Improved |
| select classical EM initialisation by likelihood instead of best Jaccard (ablation) | classical EM · DAVIS2016 val | Jaccard J: 72.3 -> 63.7 | Hurt |
| remove CRF post-processing from CIS (ablation) | CIS · DAVIS2016 val | Jaccard J: 71.5 -> 59.2 | Hurt |
Closest papers by meaning
- Unsupervised Space-Time Network for Temporally-Consistent Segmentation of Multiple Motions (CVPR 2023 · similarity 0.90)
- Segmenting the Motion Components of a Video: A Long-Term Unsupervised Model (TPAMI 2025 · similarity 0.89)
- Learning Segmentation from Point Trajectories (NeurIPS 2024 · similarity 0.88)
- UnSAMFlow: Unsupervised Optical Flow Guided by Segment Anything Model (CVPR 2024 · similarity 0.88)
- Competitive Collaboration: Joint Unsupervised Learning of Depth, Camera Motion, Optical Flow and Motion Segmentation (CVPR 2019 · similarity 0.88)
- FlowFeat: Pixel-Dense Embedding of Motion Profiles (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
