Research map / Vision-language and multimodal / CLIP and zero-shot
Hierarchy-Aware Pseudo Word Learning with Text Adaptation for Zero-Shot Composed Image Retrieval
Zhe Li, Lei Zhang, Zheren Fu, Kun Zhang, 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 object-aware tokens s_s on top of global token (ablation) | HIT (BLIP) · CIRR val | R@1: 33.2 -> 35.0 | Improved |
| add object-aware tokens s_p on top of global token (ablation) | HIT (BLIP) · CIRR val | R@1: 33.2 -> 35.7 | Improved |
| use all three levels s_g, s_s and s_p together (ablation) | HIT (BLIP) · CIRR val | R@1: 35.7 -> 37.3 | Improved |
| use all three levels instead of s_g plus s_s (ablation) | HIT (BLIP) · Fashion-IQ val | R@10: 32.9 -> 34.5 | Improved |
| add two-stage grouping modules (L=2 vs none) (ablation) | HIT (BLIP) · CIRR val | avg R@1+R@5: 49.08 -> 53.35 | Improved |
| increase grouping modules beyond 2 (L=2 to L=4) (ablation) | HIT (BLIP) · CIRR val | avg R@1+R@5: 53.35 -> 51.12 | Hurt |
| use 32 and 8 group tokens instead of 16 and 4 (ablation) | HIT (BLIP) · CIRR val | R@1: 34.6 -> 37.3 | Improved |
| increase group tokens from 32/8 to 64/32 (ablation) | HIT (BLIP) · CIRR val | R@1: 37.3 -> 34.9 | Hurt |
| increase group tokens from 32/8 to 32/16 (ablation) | HIT (BLIP) · Fashion-IQ val | R@10: 34.5 -> 33.8 | Hurt |
| set auxiliary loss weight to 0.7 instead of 0.1 (ablation) | HIT (BLIP) · CIRR val | avg R@1+R@5: 51.46 -> 53.35 | Improved |
| raise auxiliary loss weight from 0.7 to 1.0 (ablation) | HIT (BLIP) · CIRR val | avg R@1+R@5: 53.35 -> 52.97 | Hurt |
| use stronger BLIP-2 backbone instead of CLIP (ablation) | HIT · CIRR val | R@1: 28.7 -> 39.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Context-I2W — Context-I2W: Mapping Images to Context-Dependent Words for Accurate Zero-Shot Composed Image Retrieval (AAAI 2024) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Knowledge-Enhanced Dual-stream Zero-shot Composed Image Retrieval (CVPR 2024 · similarity 0.89)
- Context-I2W: Mapping Images to Context-Dependent Words for Accurate Zero-Shot Composed Image Retrieval (AAAI 2024 · similarity 0.89)
- Missing Target-Relevant Information Prediction with World Model for Accurate Zero-Shot Composed Image Retrieval (CVPR 2025 · similarity 0.88)
- Language-only Training of Zero-shot Composed Image Retrieval (CVPR 2024 · similarity 0.88)
- Image2Sentence based Asymmetrical Zero-shot Composed Image Retrieval (ICLR 2024 · similarity 0.87)
- ConText-CIR: Learning from Concepts in Text for Composed Image Retrieval (CVPR 2025 · 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.
