Research map / Video / Video segmentation and tracking
The Visual Centrifuge: Model-Free Layered Video Representations
Jean-Baptiste Alayrac, Joao Carreira, Andrew Zisserman
Check 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 |
|---|---|---|---|
| Train the predictor with permutation-invariant reconstruction loss instead of leaving it untrained | Predictor · Kinetics-600 val, blended video pairs | validation loss: 0.561 -> 0.187 | Improved |
| Add a correction module to the predictor (ablation) | Predictor, 4 output layers · Kinetics-600 val, blended video pairs | validation loss: 0.159 -> 0.133 | Improved |
| Increase the number of output layers from 2 to 4 (ablation) | Predictor · Kinetics-600 val, blended video pairs | validation loss: 0.187 -> 0.159 | Improved |
| Increase the number of output layers from 4 to 8 or 12 (ablation) | Predictor · Kinetics-600 val, blended video pairs | validation loss: 0.159 -> 0.151; 0.159 -> 0.150 | No clear effect |
| Add a recomposition-consistency loss (ablation) | Video layer separation model · Kinetics-600 blended video pairs | reconstruction performance | No clear effect |
| Add an explicit diversity loss between output layers (ablation) | Video layer separation model · Kinetics-600 blended video pairs | reconstruction performance | No clear effect |
| Unmix blended videos before action recognition | I3D action classifier · Kinetics-600 val, blended video pairs | two-label action accuracy: 22% -> 44% | Improved |
| Train on frozen videos without motion rather than normal videos (ablation) | Predictor-Corrector · Kinetics-600 val, two frozen videos | validation loss: 0.205 -> 0.165 | Improved |
| Train on frozen videos without motion rather than normal videos (ablation) | Predictor-Corrector · Kinetics-600 val, two normal videos | validation loss: 0.133 -> 0.233 | Hurt |
| Use a deeper encoder endpoint, Mixed_4f or Mixed_5c, instead of Mixed_3c (ablation) | Predictor · Kinetics-600 val, blended video pairs | validation loss: 0.214 -> 0.181; 0.214 -> 0.187 | Improved |
| Separate same-class video pairs rather than random pairs (ablation) | Predictor-Corrector · Kinetics-600, same-action-class pairs | average reconstruction loss: 0.133 -> 0.145 | Hurt |
Closest papers by meaning
- Extending Layered Models to 3D Motion (ECCV 2018 · similarity 0.88)
- Visual Dynamics: Stochastic Future Generation via Layered Cross Convolutional Networks (TPAMI 2018 · similarity 0.86)
- From Image to Video: An Empirical Study of Diffusion Representations (ICCV 2025 · similarity 0.85)
- LayerT2V: A Unified Multi-Layer Video Generation Framework (ICML 2026 · similarity 0.85)
- SIMONe: View-Invariant, Temporally-Abstracted Object Representations via Unsupervised Video Decomposition (NeurIPS 2021 · similarity 0.85)
- Generative Omnimatte: Learning to Decompose Video into Layers (CVPR 2025 · similarity 0.85)
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.
