Research map / Vision-language and multimodal / Human and embodied modeling
Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts
Check 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 |
|---|---|---|---|
| Replace CC3M with CC12M for pretraining before COCO fine-tuning | Image captioning model · nocaps val | overall CIDEr: 73.2 -> 87.4; +14.2 CIDEr points | Helped |
| Pretrain on CC3M and CC12M together rather than CC12M alone | Image captioning model · nocaps val | overall CIDEr: 87.4 -> 90.2; +2.8 | Helped |
| Use CC12M rather than CC3M for pretraining | Image captioning model · nocaps val out-of-domain | CIDEr: 65.3 -> 91.3 | Helped |
| Use CC12M rather than CC3M for pretraining | Image captioning model · nocaps val zero-shot | overall CIDEr: 29.7 -> 27.1 | Hurt |
| Use CC12M rather than CC3M for pretraining | Image captioning model · LocNar OID val | CIDEr: 35.7 -> 38.6 | Helped |
| Pretrain on CC12M before fine-tuning on CC3M | Image captioning model · CC3M dev | CIDEr: 100.9 -> 105.4 (+4.5 CIDER points) | Helped |
| Use CC12M rather than CC3M for pretraining | Image retrieval model · Flickr30K test, zero-shot | Recall@1: 35.4 -> 42.5 (+7.1% R1) | Helped |
| Use CC12M rather than CC3M for pretraining | Image retrieval model · Flickr30K test, fine-tuned | Recall@1: 52.3 -> 58.5 | Helped |
| Use CC12M rather than CC3M for pretraining | Image retrieval model · LocNar Flickr30K test | Recall@1: 61.1 -> 70.2 | Helped |
| Replace image captioning pretraining with MLM at low masking rates (ablation) | CC12M-pretrained image captioning model · nocaps val | overall CIDEr: 87.4 vs. 67.4, 76.8, and 84.0 | Hurt |
| Raise MLM masking rate to .8 instead of using image captioning pretraining (ablation) | CC12M-pretrained image captioning model · nocaps val | overall CIDEr: 87.4 vs. 88.5 overall CIDEr; out-of-domain CIDEr 91.3 vs. 91.1 | Mixed |
| Add visual-linguistic matching loss to image captioning pretraining (ablation) | CC12M image captioning model · nocaps val | overall CIDEr: 87.4 -> 87.1 | Hurt |
Closest papers by meaning
- The Neglected Tails in Vision-Language Models (CVPR 2024 · similarity 0.89)
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · similarity 0.88)
- Concept-Aware Batch Sampling Improves Language-Image Pretraining (CVPR 2026 · similarity 0.88)
- Distilling Vision-Language Models on Millions of Videos (CVPR 2024 · similarity 0.88)
- FlexCap: Describe Anything in Images in Controllable Detail (NeurIPS 2024 · similarity 0.88)
- A Picture is Worth More Than 77 Text Tokens: Evaluating CLIP-Style Models on Dense Captions (CVPR 2024 · 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.