Research map / Detection and segmentation / Weakly supervised detection and segmentation
Frozen CLIP-DINO: A Strong Backbone for Weakly Supervised Semantic Segmentation
Bingfeng Zhang, Siyue Yu, Jimin Xiao, Yunchao Wei, Yao Zhao
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 RFM CAM refinement module to the decoder (ablation) | WeCLIP · PASCAL VOC 2012 val (w/o CRF) | mIoU: 68.7 -> 74.9 | Improved |
| use 3 transformer layers in decoder instead of 1 (ablation) | WeCLIP decoder · PASCAL VOC 2012 val (w/o CRF) | mIoU: 73.2 -> 74.9 | Improved |
| use 5 decoder transformer layers instead of 3 (ablation) | WeCLIP decoder · PASCAL VOC 2012 val (w/o CRF) | mIoU: 74.9 -> 70.3 | Hurt |
| add decoder affinity A_f to static attention refinement A_s (ablation) | RFM refining map · PASCAL VOC 2012 val (w/o CRF) | mIoU: 71.8 -> 74.3 | Improved |
| add reliability filter G_e to A_s (ablation) | RFM refining map · PASCAL VOC 2012 val (w/o CRF) | mIoU: 71.8 -> 72.3 | Improved |
| add filter G_e to A_f + A_s (ablation) | RFM refining map · PASCAL VOC 2012 val (w/o CRF) | mIoU: 74.3 -> 74.9 | Improved |
| use frozen CLIP features from all 12 blocks instead of only block 12 (ablation) | WeCLIP decoder input · PASCAL VOC 2012 val (w/o CRF) | mIoU: 74.3 -> 74.9 | Improved |
| supervise A_f with online pseudo label M_p instead of argmax of prediction P (ablation) | WeCLIP affinity loss · PASCAL VOC 2012 val (w/o CRF) | mIoU: 74.6 -> 74.9 | Improved |
| supervise A_f with intersection of M_p and P instead of M_p (ablation) | WeCLIP affinity loss · PASCAL VOC 2012 val (w/o CRF) | mIoU: 74.9 -> 74.8 | No clear effect |
| add affinity loss weight lambda=0.1 versus 0 (ablation) | WeCLIP loss · PASCAL VOC 2012 val (w/o CRF) | mIoU: 73.3 -> 74.9 | Improved |
| increase lambda from 0.1 to 1.0 (ablation) | WeCLIP loss · PASCAL VOC 2012 val (w/o CRF) | mIoU: 74.9 -> 72.9 | Hurt |
| use multi-scale {0.75, 1.0} instead of single scale {1.0} at inference (ablation) | WeCLIP inference · PASCAL VOC 2012 val (w/o CRF) | mIoU: 74.0 -> 74.9 | 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.
- MCTformer — MCTformer+: Multi-Class Token Transformer for Weakly Supervised Semantic Segmentation (TPAMI 2024) · better on 1 of 1 shared result
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
- Frozen CLIP: A Strong Backbone for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.98)
- CLIP Is Also an Efficient Segmenter: A Text-Driven Approach for Weakly Supervised Semantic Segmentation (CVPR 2023 · similarity 0.91)
- Extract Free Dense Labels from CLIP (ECCV 2022 · similarity 0.89)
- CLIP-DINOiser: Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation (ECCV 2024 · similarity 0.89)
- Reliability Does Matter: An End-to-End Weakly Supervised Semantic Segmentation Approach (AAAI 2020 · similarity 0.89)
- Unveiling the Knowledge of CLIP for Training-Free Open-Vocabulary Semantic Segmentation (AAAI 2025 · 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.
