Research map / Video / Video segmentation and tracking
Broaden Your Views for Self-Supervised Video Learning
Adrià Recasens, Pauline Luc, Jean-Baptiste Alayrac, Luyu Wang, Florian Strub, Corentin Tallec, Mateusz Malinowski, Viorica Pătrăucean, Florent Altché, Michal Valko, Jean-Bastien Grill, Aäron van den Oord and 1 more
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 |
|---|---|---|---|
| lengthen broad view from 1.3s to 10s (narrow 1.3s) (ablation) | BraVe RGB+RC · HMDB51 linear (K600 ablation) | top-1 accuracy: 59.8 -> 65.2 | Improved |
| use 10s narrow view as well as 10s broad view (ablation) | BraVe RGB+RC · HMDB51 linear (K600 ablation) | top-1 accuracy: 65.2 -> 58.0 | Hurt |
| sync narrow and broad view start for visual broad view (ablation) | BraVe RGB+RC · UCF101 linear (K600 ablation) | top-1 accuracy: 91.7 -> 87.5 | Hurt |
| use random convolutions in the broad view instead of plain RGB (ablation) | BraVe · HMDB51 linear (K600 ablation) | top-1 accuracy: 63.2 -> 65.2 | Improved |
| use optical flow instead of RGB+RC in the broad view (ablation) | BraVe · HMDB51 linear (K600 ablation) | top-1 accuracy: 65.2 -> 66.4 | Mixed |
| add a second and third RGB+RC broad view (ablation) | BraVe · Kinetics-600 linear (K600 ablation) | top-1 accuracy: 69.1 -> 70.8 | Improved |
| add both RGB+RC and flow broad views (1+1) (ablation) | BraVe · HMDB51 linear (K600 ablation) | top-1 accuracy: 65.2 -> 68.5 | Improved |
| share the projector between narrow and broad views (ablation) | BraVe · HMDB51 linear (weight-sharing ablation, plain RGB broad) | top-1 accuracy: 61.3 -> 61.7 | Improved |
| share projector and predictor in addition (separate backbones) (ablation) | BraVe · HMDB51 linear (weight-sharing ablation, plain RGB broad) | top-1 accuracy: 61.7 -> 63.2 | Improved |
| share the predictor on top of shared backbone and projector (ablation) | BraVe · HMDB51 linear (weight-sharing ablation, plain RGB broad) | top-1 accuracy: 63.7 -> 62.7 | Hurt |
| sync audio broad view at the start of the narrow view instead of async (ablation) | BraVe audio · HMDB51 linear (AudioSet audio-broad ablation) | top-1 accuracy: 67.4 -> 68.2 | Improved |
| extend audio broad view from 1.3s to 5s (ablation) | BraVe audio · UCF101 linear (AudioSet) | top-1 accuracy: 92.3 -> 92.7 | Improved |
Closest papers by meaning
- Contrast and Order Representations for Video Self-Supervised Learning (ICCV 2021 · similarity 0.87)
- Self-Supervised Learning of Compressed Video Representations (ICLR 2021 · similarity 0.87)
- Self-supervised Co-Training for Video Representation Learning (NeurIPS 2020 · similarity 0.87)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.86)
- Self-Supervised Learning of Video-Induced Visual Invariances (CVPR 2020 · similarity 0.86)
- SMILE: Infusing Spatial and Motion Semantics in Masked Video Learning (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.
