Research map / Video / Action recognition
Self-Supervised Video Representation Learning via Latent Time Navigation
Read the paperCheck 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 |
|---|---|---|---|
| add time-aware latent navigation (ablation) | MoCo video representation learning · Toyota Smarthome CS | Top-1 accuracy: 65.1 -> 67.8 | Helped |
| add linear time addition without a decomposed basis (ablation) | MoCo video representation learning · Toyota Smarthome CS | Top-1 accuracy: 65.1 -> 66.0 | Helped |
| add time-driven attention along the time basis (ablation) | MoCo video representation learning · Toyota Smarthome CS | Top-1 accuracy: 65.1 -> 66.7 | Helped |
| add time-driven linear transformation along the orthogonal basis (ablation) | MoCo video representation learning · Toyota Smarthome CS | Top-1 accuracy: 65.1 -> 67.8 | Helped |
| orthogonalize the time basis (ablation) | LTN linear transformation · Toyota Smarthome CS | Top-1 accuracy: 67.3 -> 67.8 | Helped |
| add LTN time parameterization (ablation) | MoCo, ρMoCo, and ρBYOL · Toyota Smarthome CS | Top-1 accuracy: 61.8 -> 65.5; 65.1 -> 67.8; 61.7 -> 63.3 | Helped |
| set the time-basis size to M=64 rather than the other tested sizes (ablation) | LTN · Toyota Smarthome CS | Top-1 accuracy: M=16: 65.2; M=64: 67.8; M=128: 67.3; M=512: 67.6; M=1024: 67.5; M=2000: 66.9 | Mixed |
| use a 2-layer, 2048-dimensional time encoder (ablation) | LTN · Toyota Smarthome CS | Top-1 accuracy: None: 65.1; 1 layer, 128: 66.7; 1 layer, 1024: 67.3; 2 layers, 1024: 67.1; 2 layers, 2048: 67.8; 3 layers, 2048: 67.9 | Mixed |
| LTN pretraining on Kinetics-400 | R3D-50 · Kinetics-400 | Top-1 accuracy: 70.0 -> 71.3 | Helped |
| use LTN rather than the strongest listed single-RGB frozen-feature baseline | R3D-50 · UCF101, K400 pretraining, frozen features | accuracy: 90.2 -> 90.6 | Helped |
| use LTN rather than the strongest listed single-RGB frozen-feature baseline | R3D-50 · HMDB51, K400 pretraining, frozen features | accuracy: 58.7 -> 58.9 | Helped |
| use LTN rather than ρBYOL | R3D-50 · UCF101, K400 pretraining, fine-tuning | accuracy: 94.2 -> 94.5 | Helped |
Closest papers by meaning
- Frame-Wise Action Representations for Long Videos via Sequence Contrastive Learning (CVPR 2022 · similarity 0.90)
- Learning by Aligning Videos in Time (CVPR 2021 · similarity 0.89)
- Long-Short Temporal Contrastive Learning of Video Transformers (CVPR 2022 · similarity 0.89)
- Semi-Supervised Action Recognition With Temporal Contrastive Learning (CVPR 2021 · similarity 0.88)
- SSAN: Separable Self-Attention Network for Video Representation Learning (CVPR 2021 · similarity 0.88)
- Fine-Grained Temporal Contrastive Learning for Weakly-Supervised Temporal Action Localization (CVPR 2022 · similarity 0.88)
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.