Research map / Vision-language and multimodal / CLIP and zero-shot
PyramidCLIP: Hierarchical Feature Alignment for Vision-language Model Pretraining
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 |
|---|---|---|---|
| Use hierarchical peer-level and cross-level alignment, rather than CLIP's standard pairing | PyramidCLIP-ResNet50 · ImageNet zero-shot | top-1 accuracy: 55.3 -> 61.4 | Helped |
| Add global-relation cross-level alignment (ablation) | Peer-level ResNet50 · MS-COCO I2T | R@1: 28.5 -> 31.9(+3.4) | Helped |
| Add local-relation cross-level alignment (ablation) | Peer-level plus global-relation ResNet50 · MS-COCO I2T | R@1: 31.9 -> 34.6(+6.1) | Helped |
| Soften all contrastive objectives (ablation) | PyramidCLIP-ResNet50 · MS-COCO I2T | R@1: 34.6 -> 36.4(+7.9) | Helped |
| Add ROI-object-attribute peer-level alignment (ablation) | Peer-level ResNet50 · MS-COCO I2T | R@1: 28.5 -> 28.5 | No clear effect |
| Use local image crop with original caption instead of the standard global image-caption pairing (ablation) | ResNet50 peer-level alignment · ImageNet zero-shot | top-1 accuracy: 30.0 -> 30.8(+0.8) | Helped |
| Add global summary alignment to local-caption alignment (ablation) | ResNet50 peer-level alignment · ImageNet zero-shot | top-1 accuracy: 30.8 -> 32.8(+2.8) | Helped |
| Add LeFF to the ViT image encoder (ablation) | PyramidCLIP-ViT-B/32 · MS-COCO I2T | R@1: 31.4 -> 31.7(+7.4) | Helped |
| Set the number of ViT layers with LeFF to 9 (ablation) | PyramidCLIP-ViT · ImageNet zero-shot | top-1 accuracy | Helped |
Closest papers by meaning
- SmartCLIP: Modular Vision-language Alignment with Identification Guarantees (CVPR 2025 · similarity 0.91)
- SuperCLIP: CLIP with Simple Classification Supervision (NeurIPS 2025 · similarity 0.90)
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · similarity 0.90)
- DINOv2 Meets Text: A Unified Framework for Image- and Pixel-Level Vision-Language Alignment (CVPR 2025 · similarity 0.90)
- Kernel-based Unsupervised Embedding Alignment for Enhanced Visual Representation in Vision-language Models (ICML 2025 · similarity 0.89)
- DH-Set: Improving Vision-Language Alignment with Diverse and Hybrid Set-Embeddings Learning (CVPR 2025 · 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.