Research map / Video / Video-language understanding
Unmasked Teacher: Towards Training-Efficient Video Foundation Models
Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, Yu Qiao
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 |
|---|---|---|---|
| replace pixel reconstruction with unmasked token alignment (ablation) | UMT-B (ViT-B/16) · Kinetics-400 | top-1 accuracy (%): 78.8 -> 84.9 | Improved |
| add pixel reconstruction to unmasked token alignment (ablation) | UMT-B (ViT-B/16) · Kinetics-400 | top-1 accuracy (%): 84.9 -> 83.9 | Hurt |
| add masked token recovery (ablation) | UMT-B (ViT-B/16) · Kinetics-400 | top-1 accuracy (%): 84.9 -> 84.6 | Hurt |
| use semantic masking instead of random masking (ablation) | UMT-B (ViT-B/16) · Kinetics-400 | top-1 accuracy (%): 84.6 -> 84.9 | Improved |
| use semantic masking instead of random masking (ablation) | UMT-B (ViT-B/16) · Something-Something V2 | top-1 accuracy (%): 70.2 -> 70.2 | No clear effect |
| use sparse sampling instead of dense sampling (ablation) | UMT-B (ViT-B/16) · Kinetics-400 | top-1 accuracy (%): 84.0 -> 84.9 | Improved |
| remove temporal downsampling (ablation) | UMT-B (ViT-B/16) · Something-Something V2 | top-1 accuracy (%): 69.5 -> 70.2 | Improved |
| use spatiotemporal attention in student pre-training and fine-tuning (ablation) | UMT-B (ViT-B/16) · Something-Something V2 | top-1 accuracy (%): 54.0 -> 70.3 | Improved |
| use CLIP spatial teacher instead of spatiotemporal CLIP-ST teacher (ablation) | UMT-B (ViT-B/16) · Kinetics-400 | top-1 accuracy (%): 83.7 -> 85.2 | Improved |
| add mask modeling with spatiotemporal attention student (ablation) | UMT-B (ViT-B/16) · Something-Something V2 | top-1 accuracy (%): 69.1 -> 70.3 | Improved |
| use two-stage progressive pre-training instead of one stage (ablation) | UMT-B (ViT-B/16) · MSRVTT retrieval | average recall (%): 57.9 -> 67.0 | Improved |
| add unmasked teacher alignment in multimodal pre-training (ablation) | UMT-B (ViT-B/16) · MSRVTT retrieval | average recall (%): 66.9 -> 67.0 | Improved |
Closest papers by meaning
- Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization (ICML 2024 · similarity 0.86)
- Uni4D: Unifying Visual Foundation Models for 4D Modeling from a Single Video (CVPR 2025 · similarity 0.85)
- VONet: Unsupervised Video Object Learning With Parallel U-Net Attention and Object-wise Sequential VAE (ICLR 2024 · similarity 0.85)
- Asymmetric Masked Distillation for Pre-Training Small Foundation Models (CVPR 2024 · similarity 0.84)
- ViLMA: A Zero-Shot Benchmark for Linguistic and Temporal Grounding in Video-Language Models (ICLR 2024 · similarity 0.84)
- Needle In A Video Haystack: A Scalable Synthetic Evaluator for Video MLLMs (ICLR 2025 · similarity 0.84)
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.
