Research map / Detection and segmentation / Weakly supervised detection and segmentation
CLIP Is Also an Efficient Segmenter: A Text-Driven Approach for Weakly Supervised Semantic Segmentation
Read the paperCheck 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 |
|---|---|---|---|
| Add softmax to GradCAM (ablation) | CLIP-ES initial CAM generation · PASCAL VOC 2012 train | mIoU: 49.4$^*$ -> 53.3$^*$ | Helped |
| Include class-related background categories in softmax competition (ablation) | CLIP-ES initial CAM generation · PASCAL VOC 2012 train, boat and train | mIoU: 24.1 -> 46.9; 43.8 -> 57.5 | Helped |
| Use class-aware CAA refinement instead of vanilla MHSA (ablation) | CLIP-ES CAM refinement · PASCAL VOC 2012 train | mIoU: 68.2 -> 70.8 | Helped |
| Add CAA refinement to initial CAMs (ablation) | CLIP-ES · PASCAL VOC 2012 train | mIoU: 58.6 -> 70.8 | Helped |
| Add dense CRF after CAA refinement (ablation) | CLIP-ES · PASCAL VOC 2012 train | mIoU: 70.8 -> 75.0 | Helped |
| Average remaining ViT tokens instead of using the class token (ablation) | CLIP ViT-B-16 CAM generation · PASCAL VOC 2012 train | mIoU: 43.8 -> 58.6 | Helped |
| Use confidence-guided loss instead of cross entropy (ablation) | DeepLabV2 trained on CLIP-ES pseudo masks · PASCAL VOC 2012 | mIoU: 70.6 -> 71.1 | Helped |
| Use confidence-guided loss instead of cross entropy (ablation) | DeepLabV2 trained on CLIP-ES pseudo masks · MS COCO 2014 | mIoU: 45.1 -> 45.4 | Helped |
| Add synonym fusion to class text (ablation) | CLIP-ES initial CAM generation · PASCAL VOC 2012 train, person | mIoU: 43.6 -> 51.6 | Helped |
| Use sentence-level synonym fusion rather than feature-level or CAM-level fusion (ablation) | CLIP-ES · PASCAL VOC 2012 train, all classes | mIoU: 70.8; 70.8; 70.6 | No clear effect |
| Use multi-scale inference with CAA and dense CRF (ablation) | CLIP-ES CAM generation · PASCAL VOC 2012 train | mIoU: 75.0 / 74.1$^*$ | No clear effect |
| Use a lower confidence threshold for CGL (ablation) | CLIP-ES segmentation training · PASCAL VOC 2012 | mIoU: 73.7; 73.6; 73.8 | No clear effect |
Closest papers by meaning
- Exploring CLIP's Dense Knowledge for Weakly Supervised Semantic Segmentation (CVPR 2025 · similarity 0.92)
- Frozen CLIP: A Strong Backbone for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.92)
- Unveiling the Knowledge of CLIP for Training-Free Open-Vocabulary Semantic Segmentation (AAAI 2025 · similarity 0.90)
- TagCLIP: A Local-to-Global Framework to Enhance Open-Vocabulary Multi-Label Classification of CLIP without Training (AAAI 2024 · similarity 0.90)
- Weakly Supervised Semantic Segmentation for Driving Scenes (AAAI 2024 · similarity 0.89)
- From SAM to CAMs: Exploring Segment Anything Model for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.89)
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.