Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Pic2Word: Mapping Pictures to Words for Zero-Shot Composed Image Retrieval
Read the paperCheck 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 |
|---|---|---|---|
| Represent the query image as a learned pseudo-token and compose it with text in the frozen CLIP language encoder | Zero-shot composed image retrieval · COCO object composition | R10: 26.6 -> 33.4 | Helped |
| Represent the query image as a learned pseudo-token and compose it with text in the frozen CLIP language encoder | Zero-shot composed image retrieval · CIRR test | R10: 56.7 -> 65.3 | Helped |
| Represent the query image as a learned pseudo-token and compose it with text in the frozen CLIP language encoder | Zero-shot composed image retrieval · Fashion-IQ validation | R10: 19.8 -> 24.7 | Helped |
| Increase mapping network hidden dimension from 512 to 4096 (ablation) | Pic2Word · CIRR | R10: 65.4 -> 67.6 | Helped |
| Increase mapping network hidden dimension from 512 to 4096 (ablation) | Pic2Word · Fashion-IQ | R50: 43.7 -> 45.3 | Helped |
| Increase mapping network depth from 3 to 5 layers (ablation) | Pic2Word · COCO | R10: 33.4 -> 32.3 | Hurt |
| Remove nonlinear hidden layers and use a linear mapping (ablation) | Pic2Word · CIRR | R10: 65.4 -> 58.8 | Hurt |
| Train the mapping network on CC12M instead of CC3M (ablation) | Pic2Word · CIRR validation | R10: 65.4±1.3 -> 64.5±3.4 | No clear effect |
| Use only 10% of CC3M samples to train the mapping network (ablation) | Pic2Word · Composed image retrieval | retrieval performance | Hurt |
| Use 50% rather than all CC3M samples to train the mapping network (ablation) | Pic2Word · Composed image retrieval | retrieval performance | No clear effect |
| Replace the zero-shot method with a supervised Combiner trained on fewer than 1,000 triplets | Composed image retrieval · Fashion-IQ and CIRR | retrieval performance | Hurt |
| Fine-tune CLIP on CC3M before using the image-text baselines (ablation) | CLIP image-text baselines · COCO | R10: 26.6 -> 29.0 | Helped |
Closest papers by meaning
- Language-only Training of Zero-shot Composed Image Retrieval (CVPR 2024 · similarity 0.93)
- Context-I2W: Mapping Images to Context-Dependent Words for Accurate Zero-Shot Composed Image Retrieval (AAAI 2024 · similarity 0.92)
- Image2Sentence based Asymmetrical Zero-shot Composed Image Retrieval (ICLR 2024 · similarity 0.92)
- ConText-CIR: Learning from Concepts in Text for Composed Image Retrieval (CVPR 2025 · similarity 0.91)
- Generative Zero-Shot Composed Image Retrieval (CVPR 2025 · similarity 0.91)
- Reason-before-Retrieve: One-Stage Reflective Chain-of-Thoughts for Training-Free Zero-Shot Composed Image Retrieval (CVPR 2025 · similarity 0.91)
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.