Research map / Vision-language and multimodal / CLIP and zero-shot
SCLIP: Rethinking Self-Attention for Dense Vision-Language Inference
Feng Wang, Jieru Mei, Alan Yuille
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 |
|---|---|---|---|
| replace CLIP self-attention with correlative self-attention (query-query and key-key similarity) | CLIP ViT-B · 8 segmentation benchmarks avg | mIoU (%): 14.1 -> 38.2 | Improved |
| use CSA with the pretrained Wq and Wk ensemble instead of identity projection (ablation) | SCLIP · PASCAL VOC | mIoU (%): 57.5 -> 59.1 | Improved |
| learn the projection matrix from 64 training samples (ablation) | SCLIP · PASCAL VOC | mIoU (%): 59.1 -> 60.4 | No clear effect |
| use random projection matrices (n=1 to 16) instead of pretrained (ablation) | SCLIP · PASCAL VOC | mIoU (%): 59.1 -> 57.1 (n=1), 58.1 (n=16) | Hurt |
| sharpen CLIP attention via temperature instead of CSA (ablation) | CLIP · VOC21 / Ctx59 / Stuff | mIoU (%): 18.8 -> 21.7 (tau=2) on VOC21; 13.3 -> 9.5 on Ctx59 | No clear effect |
| use local attention (window 3) instead of CSA (ablation) | CLIP · VOC21 | mIoU (%): 59.1 -> 42.9 | Hurt |
| borrow early-layer attention maps instead of CSA (ablation) | CLIP · VOC21 | mIoU (%): 59.1 -> 43.0 (layer 3) | Hurt |
| use image size 336 / window 224 / stride 112 instead of prior 448/448/224 protocol (ablation) | SCLIP · VOC21 | mIoU (%): 58.9 -> 59.1 | Improved |
| increase image size to 448 with window 224 (ablation) | SCLIP · VOC21 | mIoU (%): 59.1 -> 60.4 | Improved |
| add PAMR post-processing | SCLIP · 8 segmentation benchmarks avg | mIoU (%): 38.2 -> 40.1 | 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.
- MaskCLIP — MaskCLIP: Masked Self-Distillation Advances Contrastive Language-Image Pretraining (CVPR 2023) · 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
- ResCLIP: Residual Attention for Training-free Dense Vision-language Inference (CVPR 2025 · similarity 0.93)
- Refining CLIP's Spatial Awareness: A Visual-Centric Perspective (ICLR 2025 · similarity 0.89)
- DeCLIP: Decoupled Learning for Open-Vocabulary Dense Perception (CVPR 2025 · similarity 0.89)
- Unveiling the Knowledge of CLIP for Training-Free Open-Vocabulary Semantic Segmentation (AAAI 2025 · similarity 0.89)
- From Local Details to Global Context: Advancing Vision-Language Models with Attention-Based Selection (ICML 2025 · similarity 0.89)
- Building Vision-Language Models on Solid Foundations with Masked Distillation (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.
