Research map / Detection and segmentation / Open-set and few-shot detection
Prompt Pre-Training with Twenty-Thousand Classes for Open-Vocabulary Visual Recognition
Shuhuai Ren, Aston Zhang, Yi Zhu, Shuai Zhang, Shuai Zheng, Mu Li, Alex Smola, Xu Sun
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 |
|---|---|---|---|
| remove local correction (adaptive margin) (ablation) | POMP (K=1000) · 10 cross-dataset classification avg | accuracy (%): 65.8 -> 67.0 | Improved |
| remove local correction (adaptive margin) (ablation) | POMP (K=1000) · ImageNet-21K | accuracy (%): 25.0 -> 25.3 | Improved |
| increase sampled classes K from 100 to 1000 (ablation) | POMP · 10 cross-dataset classification avg | accuracy (%): 65.5 -> 67.0 | Improved |
| use frequency-based negative sampling instead of uniform (ablation) | POMP · 10 cross-dataset classification avg | accuracy (%): 67.0 -> 66.2 | Hurt |
| use similarity-based hard negative sampling instead of uniform (ablation) | POMP · 10 cross-dataset classification avg | accuracy (%): 67.0 -> 64.2 | Hurt |
| margin m=1.5 instead of 1 at K=1000 (ablation) | POMP · 10 cross-dataset classification avg | accuracy (%): 67.0 -> 66.4 | Hurt |
| margin m=1.5 instead of 1 at K=319 (ablation) | POMP · 10 cross-dataset classification avg | accuracy (%): 66.2 -> 66.5 | Improved |
| prompt length 16 instead of 4 (ablation) | POMP · 10 cross-dataset classification avg | accuracy (%): 67.0 -> 65.0 | Hurt |
| pre-train prompt on ImageNet-21K (POMP) instead of ImageNet-1K tuning (CoOp) | CLIP ViT-B/16 · 10 cross-dataset classification avg | accuracy (%): 63.9 -> 67.0 | Improved |
| use local contrast sampling (K=128) instead of full class set when tuning on ImageNet-1K | POMP vs CoOp · ImageNet-1K 16-shot | accuracy (%): 71.9 -> 71.2 | No clear effect |
| use POMP prompt instead of hand-crafted prompts in ZSSeg | ZSSeg · ADE20K | mIoU: 20.2 -> 20.6 | Improved |
| POMP prompt instead of Detic prompt | Detic · LVIS AP_r | AP_r: 26.7 -> 26.8 | No clear effect |
Closest papers by meaning
- Weighted Multi-Prompt Learning with Description-free Large Language Model Distillation (ICLR 2025 · similarity 0.87)
- CoPL: Contextual Prompt Learning for Vision-Language Understanding (AAAI 2024 · similarity 0.87)
- ProAPO: Progressively Automatic Prompt Optimization for Visual Classification (CVPR 2025 · similarity 0.87)
- Just a Few Glances: Open-Set Visual Perception with Image Prompt Paradigm (AAAI 2025 · similarity 0.87)
- PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 2024 · similarity 0.87)
- Learning to Prompt with Text Only Supervision for Vision-Language Models (AAAI 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.
