Research map / Vision-language and multimodal / CLIP and zero-shot
Learning to Compose: Revisiting Proxy Task Design for Zero-Shot Composed Image Retrieval
Jingjing Zhang, Lei Zhang, Zheren Fu, Zhendong Mao
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 text-anchored visual aggregation to global image+text baseline (ablation) | FoCo · Ablation CIRR test (ViT-L/14) | Recall@1: 12.4 -> 25.8 | Improved |
| add context-conditioned completion on top of aggregation (ablation) | FoCo · Ablation CIRR test (ViT-L/14) | Recall@1: 25.8 -> 34.5 | Improved |
| use completion module with coarse patch masking only (ablation) | FoCo · Ablation CIRR test (ViT-L/14) | Recall@1: 12.4 -> 22.2 | Improved |
| use multi-positive cross-instance contrastive loss instead of InfoNCE (ablation) | FoCo · Ablation CIRR test (ViT-L/14) | Recall@1: 33.1 -> 34.5 | Improved |
| use cross-instance negatives instead of intra-image negatives (aggr) (ablation) | FoCo · Ablation CIRR test (ViT-L/14) | Recall@1: 31.0 -> 34.5 | Improved |
| keep aggregation loss during joint training (ablation) | FoCo · Ablation CIRR test (ViT-L/14) | Recall@1: 33.2 -> 34.5 | Improved |
| use LLM decomposition instead of rule-based extraction (ablation) | FoCo · Ablation CIRR test (ViT-L/14) | Recall@1: 30.7 -> 34.5 | Improved |
| use multiple local captions (K>1) instead of one (ablation) | FoCo · Ablation CIRR test (ViT-L/14) | Recall@1: 32.6 -> 34.5 | Improved |
| use text-guided query with global visual feature (ablation) | FoCo · Ablation CIRR test (ViT-L/14) | Recall@1: 33.5 -> 34.5 | Improved |
| use cross-attention instead of dot-product attention (ablation) | FoCo · Ablation CIRR test (ViT-L/14) | Recall@1: 33.0 -> 34.5 | Improved |
| use Transformer blocks in completion module (ablation) | FoCo · Ablation FashionIQ val (avg) | Recall@10: 27.5 -> 31.5 | Improved |
| increase local captions K beyond 4 (K>=6) (ablation) | FoCo · CIRCO test | mAP | Hurt |
Closest papers by meaning
- Context-I2W: Mapping Images to Context-Dependent Words for Accurate Zero-Shot Composed Image Retrieval (AAAI 2024 · similarity 0.92)
- Language-only Training of Zero-shot Composed Image Retrieval (CVPR 2024 · 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)
- Image2Sentence based Asymmetrical Zero-shot Composed Image Retrieval (ICLR 2024 · similarity 0.91)
- Vision-by-Language for Training-Free Compositional Image Retrieval (ICLR 2024 · 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.
