Research map / Vision-language and multimodal / CLIP and zero-shot
Perceptual Grouping in Contrastive Vision-Language Models
Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, Jonathon Shlens
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 class token with max pooling for spatial aggregation (ablation) | CLIPpy · PASCAL VOC val, CC-12M | mIoU: 4.0 -> 50.8 | Improved |
| replace average pooling with max pooling (ablation) | CLIPpy · PASCAL VOC val, CC-12M | mIoU: 11.6 -> 50.8 | Improved |
| replace class token or average pooling with max pooling (ablation) | CLIPpy · ImageNet zero-shot, CC-12M | top-1 accuracy: 46.0 -> 42.3 | Hurt |
| initialize image encoder with DINO instead of supervised ImageNet-1K (ablation) | CLIPpy · PASCAL VOC val, CC-12M | mIoU: 22.5 -> 50.8 | Improved |
| initialize image encoder with DINO instead of random (ablation) | CLIPpy · PASCAL VOC val, CC-12M | mIoU: 23.5 -> 50.8 | Improved |
| add Sentence-T5 text initialization (ablation) | CLIPpy · PASCAL VOC val, CC-12M | mIoU: 44.3 -> 50.8 | Improved |
| freeze the pretrained image backbone (ablation) | CLIPpy · PASCAL VOC val, CC-12M | mIoU: 50.8 -> 10.4 | Hurt |
| freeze the pretrained text backbone (ablation) | CLIPpy · PASCAL VOC val, CC-12M | mIoU: 50.8 -> 29.5 | Hurt |
| add visual token sub-sampling during training (ablation) | CLIPpy · PASCAL VOC val | mIoU: 50.9 -> 51.8 | Improved |
| apply CLIPpy modifications versus CLIP baseline | CLIP ViT-B/16 on CC-12M · ImageNet zero-shot | top-1 accuracy: 46.0 -> 45.3 | No clear effect |
| apply CLIPpy modifications versus CLIP baseline | CLIP ViT-B/16 on CC-12M · ADE20K and COCO bottom-up grouping | Jaccard Similarity: 22.9 -> 28.9 (ADE20K), 20.4 -> 26.0 (COCO) | Improved |
| use CLIPpy grouping rather than baseline CLIP for zero-shot Waterbirds | CLIP · Waterbirds test, waterbird accuracy on incommensurate background | accuracy drop (domain gap): CLIP 80.2 -> 48.1; CLIPpy 76.9 -> 74.9 | Improved |
Closest papers by meaning
- Grounding Everything: Emerging Localization Properties in Vision-Language Transformers (CVPR 2024 · similarity 0.89)
- What's in the Image? A Deep-Dive into the Vision of Vision Language Models (CVPR 2025 · similarity 0.89)
- Synthesize Diagnose and Optimize: Towards Fine-Grained Vision-Language Understanding (CVPR 2024 · similarity 0.89)
- Mechanisms of Object Localization in Vision-Language Models (CVPR 2026 · similarity 0.88)
- Two Effects, One Trigger: On the Modality Gap, Object Bias, and Information Imbalance in Contrastive Vision-Language Models (ICLR 2025 · similarity 0.88)
- Not Only Text: Exploring Compositionality of Visual Representations in Vision-Language Models (CVPR 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.
