Research map / Video / Video-language understanding
HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips
Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, Josef Sivic
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 |
|---|---|---|---|
| add intra-video negative pairs (ablation) | Ours · YouCook2 clip retrieval | R@10 (%): 18.1 -> 24.8 | Improved |
| add intra-video negative pairs (ablation) | Ours · CrossTask step localization | Average recall (%): 25.7 -> 33.6 | Improved |
| add intra-video negative pairs (ablation) | Ours · LSMDC clip retrieval | R@10 (%): 12.3 -> 14.0 | Improved |
| add intra-video negative pairs (ablation) | Ours · MSR-VTT clip retrieval | R@10 (%): 30.1 -> 29.6 | Hurt |
| pre-train on HowTo100M then fine-tune on YouCook2 instead of training on YouCook2 only | Ours · YouCook2 clip retrieval | R@10 (%): 21.5 -> 35.3 | Improved |
| train on HowTo100M instead of YouCook2 only | Ours · YouCook2 clip retrieval | R@10 (%): 21.5 -> 24.8 | Improved |
| train on HowTo100M instead of MSR-VTT only | Ours · MSR-VTT clip retrieval | R@10 (%): 48.0 -> 29.6 | Hurt |
| pre-train on HowTo100M then fine-tune on MSR-VTT instead of training on MSR-VTT only | Ours · MSR-VTT clip retrieval | R@10 (%): 48.0 -> 52.8 | Improved |
| pre-train on HowTo100M then fine-tune on LSMDC instead of training on LSMDC only | Ours · LSMDC clip retrieval | R@10 (%): 25.0 -> 27.9 | Improved |
| pre-train on HowTo100M then fine-tune on LSMDC instead of training on LSMDC only | Ours · LSMDC clip retrieval | R@1 (%): 7.2 -> 7.1 | No clear effect |
| increase training set size of HowTo100M (top 2 to top 200 rank threshold) (ablation) | Ours · CrossTask, LSMDC, MSR-VTT, YouCook2 | average recall / R@10 | Improved |
| use only top-scoring positive pairs (max pool rate 0.2 instead of 1.0) (ablation) | Ours · MSR-VTT clip retrieval | R@10 (%): 29.6 -> 21.9 | Hurt |
Closest papers by meaning
- VideoUFO: A Million-Scale User-Focused Dataset for Text-to-Video Generation (NeurIPS 2025 · similarity 0.88)
- Panda-70M: Captioning 70M Videos with Multiple Cross-Modality Teachers (CVPR 2024 · similarity 0.88)
- InternVid: A Large-scale Video-Text Dataset for Multimodal Understanding and Generation (ICLR 2024 · similarity 0.88)
- Mimir: Improving Video Diffusion Models for Precise Text Understanding (CVPR 2025 · similarity 0.87)
- TOPA: Extending Large Language Models for Video Understanding via Text-Only Pre-Alignment (NeurIPS 2024 · similarity 0.87)
- Distilling Vision-Language Models on Millions of Videos (CVPR 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.
