Research map / Video / Action recognition
TimeBalance: Temporally-Invariant and Temporally-Distinctive Video Representations for Semi-Supervised Action Recognition
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 the temporally-distinctive teacher to the student and invariant-teacher setup (ablation) | TimeBalance · UCF101, 5% labels | Top-1 accuracy: 47.95 -> 52.14 | Helped |
| Add the temporally-distinctive teacher to the student and invariant-teacher setup (ablation) | TimeBalance · UCF101, 20% labels | Top-1 accuracy: 79.28 -> 82.02 | Helped |
| Add the temporally-invariant teacher to the student and distinctive-teacher setup (ablation) | TimeBalance · UCF101, 5% labels | Top-1 accuracy: 48.81 -> 52.14 | Helped |
| Add temporal-similarity-based teacher reweighting instead of uniform averaging (ablation) | TimeBalance · UCF101, 5% labels | Top-1 accuracy: 52.14 -> 53.48 | Helped |
| Add temporal-similarity-based teacher reweighting instead of uniform averaging (ablation) | TimeBalance · UCF101, 20% labels | Top-1 accuracy: 82.02 -> 83.26 | Helped |
| Pair teachers with different temporal objectives rather than two invariant teachers (ablation) | TimeBalance · UCF101, 5% labels | Top-1 accuracy: 48.33 -> 52.14 | Helped |
| Pair teachers with different temporal objectives rather than two distinctive teachers (ablation) | TimeBalance · UCF101, 20% labels | Top-1 accuracy: 80.49 -> 82.02 | Helped |
| Use four clips rather than two for distinctive-teacher pretraining (ablation) | TimeBalance distinctive teacher · UCF101, 5% labels | Top-1 accuracy: 39.55 -> 44.08 | Helped |
| Use four clips rather than two for invariant-teacher pretraining (ablation) | TimeBalance invariant teacher · UCF101, 5% labels | Top-1 accuracy: 42.78 -> 43.79 | Helped |
| Use L2 rather than KL-divergence for teacher distillation (ablation) | TimeBalance · UCF101, 5% labels | Top-1 accuracy: 52.62 -> 53.48 | Helped |
| Use L2 rather than JS-divergence for teacher distillation (ablation) | TimeBalance · UCF101, 5% labels | Top-1 accuracy: 50.91 -> 53.48 | Helped |
| Train the student from random initialization rather than video SSL initialization (ablation) | TimeBalance · UCF101, 5% labels | Top-1 accuracy: 53.5 -> 53.1 | No clear effect |
Closest papers by meaning
- Semi-Supervised Action Recognition With Temporal Contrastive Learning (CVPR 2021 · similarity 0.93)
- Stable Mean Teacher for Semi-supervised Video Action Detection (AAAI 2025 · similarity 0.90)
- SeFAR: Semi-supervised Fine-grained Action Recognition with Temporal Perturbation and Learning Stabilization (AAAI 2025 · similarity 0.87)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.87)
- Chirality in Action: Time-Aware Video Representation Learning by Latent Straightening (NeurIPS 2025 · similarity 0.87)
- Temporal Alignment-Free Video Matching for Few-shot Action Recognition (CVPR 2025 · similarity 0.87)
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.