Research map / Video / Video segmentation and tracking
Contextualized Spatio-Temporal Contrastive Learning With Self-Supervision
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Pretrain with contextualized region prediction and jointly learn holistic and local features | R3D50 video representations · AVA-Kinetics and AVA | mAP: 39.4 vs. 31.6; 31.1 vs. 24.9 | Helped |
| Pretrain with ConST-CL rather than CVRL | R3D50 video representations · OTB2015 | Precision: 78.1 vs. 75.4 | Helped |
| Pretrain with ConST-CL rather than CVRL | video action recognition representations · UCF101 (fine-tuning) | Top-1 accuracy: 92.2 -> 94.8 | Helped |
| Pretrain with ConST-CL rather than CVRL | video action recognition representations · UCF101 (linear) | Top-1 accuracy: 89.2 vs. 89.1 | No clear effect |
| Sample center or nearest temporal frames rather than random frames (ablation) | ConST-CL · UCF101, HMDB, AVA-Kinetics, and OTB | downstream task scores: UCF: 56.2 -> 94.1; HMDB: 57.8 -> 67.7; AVA-K: 34.8 -> 36.9; OTB: 75.4 -> 77.1 | Helped |
| Branch res5 and attach global and local losses to C5_g and C5_r (ablation) | ConST-CL · UCF101, HMDB, AVA-Kinetics, and OTB | downstream task scores: UCF: 93.5 -> 94.3; HMDB: 67.5 -> 68.7; AVA-K: 32.0 -> 36.7; OTB: 75.3 -> 77.7 | Helped |
| Add target-view context to region prediction (ablation) | ConST-CL · AVA-Kinetics (GT boxes) | mAP: 35.3 -> 37.6 | Helped |
| Add target-view context to region prediction (ablation) | ConST-CL · OTB2015 | Precision: 75.4 -> 79.0 | Helped |
| Use random temporal sampling rather than center sampling (ablation) | ConST-CL · UCF101 | Top-1 accuracy: 94.1 -> 56.2 | Hurt |
| Use different region-box sources, including SLIC, graph segmentation, or detectors, rather than random boxes (ablation) | ConST-CL · UCF101, AVA-Kinetics, and OTB | downstream task scores: UCF: 94.3, 93.4, 94.1, 93.4, 93.7; AVA-K: 36.9, 36.4, 36.9, 36.7, 37.2 | No clear effect |
Closest papers by meaning
- Contrastive Spatio-Temporal Pretext Learning for Self-Supervised Video Representation (AAAI 2022 · similarity 0.90)
- Spatiotemporal Contrastive Video Representation Learning (CVPR 2021 · similarity 0.90)
- Revisiting Contrastive Methods for Unsupervised Learning of Visual Representations (NeurIPS 2021 · similarity 0.89)
- When Does Contrastive Visual Representation Learning Work? (CVPR 2022 · similarity 0.88)
- Self-Supervised Spatio-Temporal Representation Learning for Videos by Predicting Motion and Appearance Statistics (CVPR 2019 · similarity 0.87)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · 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.