Research map / Vision-language and multimodal / CLIP and zero-shot
Concept Regions Matter: Benchmarking CLIP with a New Cluster-Importance Approach
Aishwarya Agarwal, Srikrishna Karanam, Vineet Gandhi
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 cluster-based concept masking (CCI) instead of Grad-ECLIP for attribution | CLIP ViT-B/16 · ImageNet-1K val | Deletion AUC Top-1 (lower better): 0.2464 -> 0.1809 | Improved |
| use CCI instead of Grad-ECLIP on COCO image retrieval | CLIP ViT-B/16 · MS COCO IR | Deletion AUC @5 (lower better): 0.2670 -> 0.1056 | Improved |
| use CCI instead of Grad-ECLIP for insertion | CLIP ViT-B/16 · ImageNet-1K val | Insertion AUC Top-5: 0.5993 -> 0.6518 | Improved |
| add Grad-ECLIP lambda_i weighting (ablation) | Grad-ECLIP · ImageNet-1K val | Deletion AUC Top-1 (lower better): 0.2535 -> 0.2464 | Improved |
| vary background with synthetic editing (COVAR Bg-varied) | CLIP ViT-B/16 · COVAR | BG-Er (%): 6.7 -> 15.6 | Hurt |
| reduce object scale | CLIP variants · COVAR | accuracy: 62.2 -> 32.6 (ViT-L/14 DataComp-1B) | Hurt |
| reduce object scale | ViT-H-qgelu 378px · COVAR | accuracy: 65.2 -> 39.3 | Hurt |
| crop, horizontal flip or translation of images | CLIP variants · COVAR | accuracy: 55.8 -> 55.2 (H-flip), 55.8 (Translate), 59.4 (Crop), ViT-B/32 DataComp-1B | No clear effect |
| use smaller patch size (B/16 vs B/32) with DataComp-1B (ablation) | CLIP ViT-B · COVAR Scale | BG-Er (%): 50.7 -> 30.5 | Improved |
| raise input resolution to 512px (ablation) | ViT-B-SigLIP2 · COVAR | BG-Er under scale: 39.3% BG-Er under scale; Bg-varied accuracy 64.9 | No clear effect |
| scale up model size | CLIP · COVAR Scale | BG-Er (%): 32.7 (ViT-bigG) | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Grad-ECLIP — Grad-ECLIP: Gradient-Based Visual and Textual Explanations for CLIP (TPAMI 2026) · better on 8 of 8 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
- Interpreting and Analysing CLIP's Zero-Shot Image Classification via Mutual Knowledge (NeurIPS 2024 · similarity 0.88)
- Demystifying CLIP Data (ICLR 2024 · similarity 0.88)
- Calibrating Multi-modal Representations: A Pursuit of Group Robustness without Annotations (CVPR 2024 · similarity 0.87)
- PerceptionCLIP: Visual Classification by Inferring and Conditioning on Contexts (ICLR 2024 · similarity 0.87)
- Interpreting CLIP with Sparse Linear Concept Embeddings (SpLiCE) (NeurIPS 2024 · similarity 0.87)
- CLIP Under the Microscope: A Fine-Grained Analysis of Multi-Object Representation (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.
