Research map / Detection and segmentation / Open-set and few-shot detection
Weak-to-Strong Compositional Learning from Generative Models for Language-based Object Detection
Kwanyong Park, Kuniaki Saito, Donghyun Kim
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 weak-to-strong compositional learning with synthetic triplets | GLIP-T · OmniLabel | AP: 19.3 -> 24.3 | Improved |
| add weak-to-strong compositional learning with synthetic triplets | FIBER-B · OmniLabel | AP: 25.7 -> 30.5 | Improved |
| add our method on top of text-augmentation DesCo | Desco-GLIP · OmniLabel | AP: 23.8 -> 26.5 | Improved |
| naively finetune on generated triplets only (ablation) | FIBER-B · OmniLabel | AP-c: 30.3 -> 27.7 | Hurt |
| add detection data and freeze visual backbone (ablation) | FIBER-B · OmniLabel | AP: 25.5 -> 26.8 | Improved |
| add intra-class negatives (description-aware contrastive learning) (ablation) | FIBER-B · OmniLabel | AP-d: 23.4 -> 27.4 | Improved |
| add structural negatives alone (ablation) | FIBER-B · OmniLabel | AP: 29.0 -> 29.0 | No clear effect |
| add structural positives (ablation) | FIBER-B · OmniLabel | AP-dL: 16.2 -> 21.3 | Improved |
| use weak-to-strong pseudo boxes instead of grounding-based (ablation) | FIBER-B · OmniLabel | AP: 29.3 -> 30.5 | Improved |
| increase descriptions per entity from 5 to 20 (ablation) | FIBER-B · OmniLabel | AP: 29.1 -> 30.5 | Improved |
| increase images per description from 2 to 8 (ablation) | FIBER-B · OmniLabel | AP-dL: 19.3 -> 21.3 | Improved |
| increase entity diversity from COCO (80) to LVIS (1203) (ablation) | FIBER-B · OmniLabel | AP-d: AP-d 28.5 -> 31.3; AP-c 31.0 -> 30.9 | Mixed |
Closest papers by meaning
- Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data (CVPR 2025 · similarity 0.90)
- TripletCLIP: Improving Compositional Reasoning of CLIP via Synthetic Vision-Language Negatives (NeurIPS 2024 · similarity 0.89)
- Contrasting Intra-Modal and Ranking Cross-Modal Hard Negatives to Enhance Visio-Linguistic Compositional Understanding (CVPR 2024 · similarity 0.89)
- Not Only Text: Exploring Compositionality of Visual Representations in Vision-Language Models (CVPR 2025 · similarity 0.88)
- Unified Generative and Discriminative Training for Multi-modal Large Language Models (NeurIPS 2024 · similarity 0.88)
- Generating Enhanced Negatives for Training Language-Based Object Detectors (CVPR 2024 · similarity 0.88)
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.
