Research map / Vision-language and multimodal / CLIP and zero-shot
Learning Visual Representations with Caption Annotations
Mert Bulent Sariyildiz, Julien Perez, Diane Larlus
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 5K nouns+adjectives+verbs from captions as tags instead of 80 ground-truth labels | TP (ResNet50, COCO) · VOC07 linear SVM | mAP: 80.2 -> 82.6 | Improved |
| use 5K caption tags instead of 80 ground-truth labels | TP (ResNet50, COCO) · IN-1K logistic regression | top-1 accuracy: 34.0 -> 43.9 | Improved |
| add adjectives to noun tags (ablation) | TP_postag (ResNet50) · VOC07 linear SVM | mAP: 81.8 -> 82.3 | Improved |
| increase tag vocabulary from 5K to 10K (NN+ADJ+VB) (ablation) | TP_postag (ResNet50) · VOC07 linear SVM | mAP: 82.6 -> 81.9 | Hurt |
| replace topic-model tags (LDA) with caption-cluster tags | TP (VGG16, COCO) · VOC07 linear probe, C-13 | mAP: 76.3 -> 77.2 | Improved |
| use clustering of captions instead of individual noun/adjective/verb tokens | TP (VGG16, COCO) · VOC07 linear probe, C-13 | mAP: 77.2 -> 79.4 | No clear effect |
| add masked language modeling with ICMLM_tfm on top of tag prediction | TP_postag (VGG16, COCO) · VOC07 linear probe, C-13 | mAP: 79.4 -> 80.5 | Improved |
| add masked language modeling with ICMLM on top of tag prediction | TP_postag (ResNet50, COCO) · IN-1K logistic regression | top-1 accuracy: 43.9 -> 51.9 | Improved |
| use ICMLM_att-fc instead of ICMLM_tfm | ICMLM (VGG16, VG) · VOC07 linear probe, C-13 | mAP: 82.6 -> 85.0 | Improved |
| use ICMLM_att-fc instead of ICMLM_tfm | ICMLM (VGG16, VG) · IN-1K linear probe, C-13 | top-1 accuracy: 51.1 -> 47.8 | Hurt |
| use 2 hidden layers and 12 heads instead of 1 layer and 1 head (ablation) | ICMLM_tfm (ResNet50) · VOC07 linear SVM | mAP: 85.7 -> 84.8 | Hurt |
| use 2 hidden layers instead of 1 in ICMLM_tfm (ablation) | ICMLM_tfm (ResNet50) · IN-1K logistic regression | top-1 accuracy: 50.6 -> 46.6 | Hurt |
Closest papers by meaning
- Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision (ICML 2021 · similarity 0.91)
- From Show to Tell: A Survey on Deep Learning-Based Image Captioning (TPAMI 2022 · similarity 0.91)
- Learning Transferable Visual Models From Natural Language Supervision (ICML 2021 · similarity 0.90)
- Linearly Mapping from Image to Text Space (ICLR 2023 · similarity 0.90)
- Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · similarity 0.89)
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.
