Research map / Label-efficient and robust learning / Recognition and long-tail
From Red Wine to Red Tomato: Composition With Context
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 |
|---|---|---|---|
| compose attribute and object classifiers with the learned transform | visual classifier composition · MITStates unseen attribute-object pairs | AP: 10.4 vs 8.8; 10.4 vs 7.9 | Helped |
| compose subject, predicate, and object classifiers with the learned transform | visual classifier composition · StanfordVRD unseen subject-predicate-object tuples | AP: 5.7 vs 4.9; 5.7 vs 4.3 | Helped |
| replace cross-entropy loss with regression loss (ablation) | composition transform on MITStates · MITStates unseen attribute-object pairs | AP: 10.4 vs 3.1 | Hurt |
| combine cross-entropy and regression losses (ablation) | composition transform on MITStates · MITStates unseen attribute-object pairs | AP: 10.4 vs 7.6 | Hurt |
| use identity initialization instead of Gaussian initialization (ablation) | composition transform on MITStates · MITStates unseen attribute-object pairs | AP: 9.8 -> 10.4 | Helped |
| use identity initialization instead of Xavier initialization (ablation) | composition transform on MITStates · MITStates unseen attribute-object pairs | AP: 9.9 -> 10.4 | Helped |
| increase the unseen-pair ratio from 0.1 to 0.5 (ablation) | composition transform on MITStates · MITStates unseen attribute-object pairs | AP: 29.8 -> 8.2 | Hurt |
| compose using entirely unseen attribute and object primitives (ablation) | composition transform on MITStates · MITStates pairs formed from unseen attributes and objects | AP: 9.6 vs 8.4 | Helped |
| use the transformation instead of visual late fusion | composition transform on MITStates · MITStates unseen attribute-object pairs | Top-3 accuracy: 27.6 vs 20.6 | Helped |
| use the transformation instead of label embeddings | composition transform on StanfordVRD · StanfordVRD unseen subject-predicate-object tuples | Top-3 accuracy: 12.7 vs 10.6 | Helped |
Closest papers by meaning
- Does Data Scaling Lead to Visual Compositional Generalization? (ICML 2025 · similarity 0.87)
- Towards Compositionality in Concept Learning (ICML 2024 · similarity 0.87)
- Learning to Compose: Improving Object Centric Learning by Injecting Compositionality (ICLR 2024 · similarity 0.86)
- ConceptMix: A Compositional Image Generation Benchmark with Controllable Difficulty (NeurIPS 2024 · similarity 0.85)
- When does compositional structure yield compositional generalization? A kernel theory. (ICLR 2025 · similarity 0.85)
- Context-based and Diversity-driven Specificity in Compositional Zero-Shot Learning (CVPR 2024 · similarity 0.84)
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.