Research map / Video / Video-language understanding
Frozen CLIP Models Are Efficient Video Learners
Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, Hongsheng Li
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 |
|---|---|---|---|
| use 4 decoder blocks instead of 1 (ablation) | EVL ViT-B/16 · Kinetics-400 | top-1 accuracy: 81.1 -> 82.9 | Improved |
| feed decoder blocks multi-layer intermediate features [-4,-3,-2,-1] instead of last layer only (ablation) | EVL ViT-B/16 · Kinetics-400 | top-1 accuracy: 82.1 -> 82.9 | Improved |
| use spatiotemporal unpooled features instead of temporal-only token features (ablation) | EVL ViT-B/16 · Kinetics-400 | top-1 accuracy: 79.8 -> 82.9 | Improved |
| use spatiotemporal features instead of spatial-only averaged features (ablation) | EVL ViT-B/16 · Kinetics-400 | top-1 accuracy: 80.1 -> 82.9 | Improved |
| add temporal modules (T-Conv, T-PE, T-CA) (ablation) | EVL ViT-B/16 · Kinetics-400 | top-1 accuracy: 82.5 -> 82.9 | No clear effect |
| add temporal modules (T-Conv, T-PE, T-CA) (ablation) | EVL ViT-B/16 · Something-Something-v2 | top-1 accuracy: 47.2 -> 61.0 | Improved |
| use 12 decoder blocks instead of 4 (ablation) | EVL ViT-B/16 · Something-Something-v2 | top-1 accuracy: 58.6 -> 61.0 | Improved |
| increase decoder depth from 4 to 5 (ablation) | EVL ViT-B/16 · Kinetics-400 | top-1 accuracy: 82.9 -> 83.0 | No clear effect |
| freeze CLIP backbone vs fully finetune (CLIP pretrain) | EVL vs TimeSformer DividedST · Kinetics-400 | top-1 accuracy: 82.0 -> 82.9 | Improved |
| freeze backbone vs fully finetune (ImageNet-21k pretrain) | EVL vs TimeSformer DividedST · Kinetics-400 | top-1 accuracy: 78.0 -> 75.4 | Hurt |
| ensemble EVL ViT-B/16 (8) with Uniformer-B (32) | Uniformer-B · Kinetics-400 | top-1 accuracy: 82.9 -> 85.2 | Improved |
| ensemble EVL (32) with Uniformer-B (32) | Uniformer-B · Something-Something-v2 | top-1 accuracy: 71.2 -> 72.1 | Improved |
Closest papers by meaning
- Beyond Short Clips: End-to-End Video-Level Learning With Collaborative Memories (CVPR 2021 · similarity 0.89)
- LookWhen? Fast Video Recognition by Learning When, Where, and What to Compute (NeurIPS 2026 · similarity 0.87)
- ViLT-CLIP: Video and Language Tuning CLIP with Multimodal Prompt Learning and Scenario-Guided Optimization (AAAI 2024 · similarity 0.87)
- Demystifying CLIP Data (ICLR 2024 · similarity 0.87)
- A Hard-to-Beat Baseline for Training-free CLIP-based Adaptation (ICLR 2024 · similarity 0.87)
- Frozen Transformers in Language Models Are Effective Visual Encoder Layers (ICLR 2024 · 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.
