Research map / Detection and segmentation / Weakly supervised detection and segmentation
What is Where by Looking: Weakly-Supervised Open-World Phrase-Grounding without Text Inputs
Tal Shaharabany, Yoad Tewel, Lior Wolf
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 |
|---|---|---|---|
| Use the proposed weakly supervised localization network instead of the strongest listed baseline | CUB with VGG16 · CUB-200-2011 | localization accuracy: 91.07 -> 93.94 | Improved |
| Use the proposed network instead of FG | Stanford-Cars · Stanford-Cars | localization accuracy: 96.2 -> 98.9 | Improved |
| Use the proposed network instead of FG | Stanford-Dogs · Stanford-Dogs | localization accuracy: 79.2 -> 86.4 | Improved |
| Use the proposed phrase-grounding network instead of GAE | VG-trained phrase grounding on VG · VG | pointing game accuracy: 54.72 -> 62.31 | Improved |
| Use Selective Search proposals and the complete WWbL inference method instead of GAE heatmaps | WWbL · Flickr30K, MS-COCO trained | point accuracy: 56.25 -> 61.38 | Improved |
| Add the CLIP relevancy-map loss to the other three losses (ablation) | WSOL network on CUB · CUB | localization accuracy: 90.02 -> 96.54 | Improved |
| Add regularization to the relevancy-map, foreground, and background losses (ablation) | WSOL network on CUB · CUB | localization accuracy: 91.08 -> 96.54 | Improved |
| Add background loss to the relevancy-map, foreground, and regularization losses (ablation) | WSOL network on CUB · CUB | localization accuracy: 94.73 -> 96.54 | Improved |
| Add foreground loss to the relevancy-map, background, and regularization losses (ablation) | WSOL network on CUB · CUB | localization accuracy: 92.67 -> 96.54 | Improved |
| Train with only the CLIP relevancy-map loss, removing the other losses (ablation) | WSOL network on CUB · CUB | localization accuracy: 96.54 -> 87.02 | Hurt |
| Set the relevancy-map loss weight lambda3 to 4 rather than 0 (ablation) | WSOL network on CUB · CUB | localization accuracy: 90.02 -> 96.50 | Improved |
| Replace Selective Search with iterative region removal (ablation) | WWbL inference · VG, VG trained | pointing game accuracy: 43.91 -> 35.02 | Hurt |
Closest papers by meaning
- Align2Ground: Weakly Supervised Phrase Grounding Guided by Image-Caption Alignment (ICCV 2019 · similarity 0.92)
- Contrastive Learning for Weakly Supervised Phrase Grounding (ECCV 2020 · similarity 0.91)
- Phrase Localization Without Paired Training Examples (ICCV 2019 · similarity 0.90)
- Improved Visual Grounding through Self-Consistent Explanations (CVPR 2024 · similarity 0.88)
- Momentum Pseudo-Labeling for Weakly Supervised Phrase Grounding (AAAI 2025 · similarity 0.88)
- Your Large Vision-Language Model Only Needs A Few Attention Heads For Visual Grounding (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.
