Research map / Detection and segmentation / Open-set and few-shot detection
Convolutions Die Hard: Open-Vocabulary Segmentation with Single Frozen Convolutional CLIP
Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, Liang-Chieh Chen
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 frozen CLIP for all three modules instead of trainable CLIP (ablation) | FC-CLIP · ADE20K zero-shot panoptic | PQ: 24.1 -> 26.8 | Improved |
| Add frozen out-of-vocabulary classifier to trainable mask generator and in-vocab classifier (ablation) | FC-CLIP · ADE20K zero-shot panoptic | PQ: 17.7 -> 25.4 | Improved |
| Trainable in-vocab classifier versus frozen out-of-vocab classifier on unseen classes (ablation) | FC-CLIP · ADE20K | PQ unseen: 12.6 -> 2.6 | Hurt |
| Share single frozen backbone instead of trainable plus frozen CLIP copies (ablation) | FC-CLIP · ADE20K zero-shot panoptic | PQ: 25.4 -> 26.8 | Mixed |
| Geometric instead of arithmetic ensemble (ablation) | FC-CLIP · ADE20K zero-shot panoptic | PQ: 25.6 -> 26.8 | Improved |
| Use alpha=0.4, beta=0.8 instead of in-vocab only (0.0, 0.0) (ablation) | FC-CLIP · ADE20K zero-shot panoptic | PQ: 17.8 -> 26.8 | Improved |
| Use out-of-vocab only (alpha=1, beta=1) instead of tuned ensemble (ablation) | FC-CLIP · ADE20K zero-shot panoptic | PQ: 26.8 -> 21.9 | Hurt |
| Swap seen/unseen weights (alpha=1.0, beta=0.0) (ablation) | FC-CLIP · ADE20K zero-shot panoptic | PQ: 26.8 -> 17.5 | Hurt |
| Increase CLIP input resolution from 224 to 1120 with ConvNeXt-L (ablation) | CLIP classifier · COCO | PQ: 17.3 -> 29.3 | Improved |
| Increase CLIP input resolution from 224 to 1120 with ViT-L/14 (ablation) | CLIP classifier · COCO | PQ: 19.3 -> 14.9 | Hurt |
| Increase ImageNet input resolution from 224 to 896 with ViT-L/14 (ablation) | CLIP classifier · ImageNet-1k | accuracy: 75.3 -> 53.9 | Hurt |
| Increase ImageNet input resolution from 224 to 336 with ConvNeXt-L (ablation) | CLIP classifier · ImageNet-1k | accuracy: 75.1 -> 77.1 | 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.
- FreeSeg — FreeSeg: Unified, Universal and Open-Vocabulary Image Segmentation (CVPR 2023) · better on 3 of 3 shared results
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
- Open-Vocabulary Semantic Segmentation With Mask-Adapted CLIP (CVPR 2023 · similarity 0.93)
- A Simple Baseline for Open-Vocabulary Semantic Segmentation with Pre-trained Vision-Language Model (ECCV 2022 · similarity 0.93)
- Unveiling the Knowledge of CLIP for Training-Free Open-Vocabulary Semantic Segmentation (AAAI 2025 · similarity 0.91)
- Open-Vocabulary Segmentation with Semantic-Assisted Calibration (CVPR 2024 · similarity 0.91)
- Mask-Adapter: The Devil is in the Masks for Open-Vocabulary Segmentation (CVPR 2025 · similarity 0.90)
- Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models (CVPR 2024 · similarity 0.90)
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.
