Research map / Detection and segmentation / Open-set and few-shot detection
A Semantic Space is Worth 256 Language Descriptions: Make Stronger Segmentation Models with Descriptive Properties
Junfei Xiao, Ziqi Zhou, Wenxuan Li, Shiyi Lan, Jieru Mei, Zhiding Yu, Bingchen Zhao, Alan Yuille, Yuyin Zhou, Cihang Xie
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 category-level label space with property-level label space | ViT-Adapter DeiT-B · Pascal Context | mIoU: 53.3 -> 58.2 | Improved |
| replace category-level label space with property-level label space | ViT-Adapter DeiT-B · BDD | mIoU: 60.7 -> 65.7 | Improved |
| replace category-level label space with property-level label space | ViT-Adapter DeiT-B · ADE20K | mIoU: 48.4 -> 49.0 | Improved |
| add ProLab property-level label space to DeepLabv3+ and SegFormer | DeepLabv3+ ResNet50 / SegFormer MiT-B1 · ADE20K, Cityscapes, BDD | mIoU: DeepLabv3+ ADE20K 42.7 -> 43.6; SegFormer ADE20K 41.5 -> 42.3 | Improved |
| cluster description embeddings into properties (vs no clustering) (ablation) | ProLab · ADE20K | mIoU: 30.2 -> 48.3 | Improved |
| vary number of clusters (64, 128, 256, 512) (ablation) | ProLab · ADE20K | mIoU: 47.8 / 48.0 / 48.3 / 47.6 | No clear effect |
| use property-level embedding space instead of category-level embedding space (BGE-B) (ablation) | ProLab · ADE20K | mIoU: 47.7 -> 49.0 | Improved |
| use BERT or T5 text encoders instead of sentence embedding model (ablation) | ProLab · ADE20K | mIoU: 47.7 (BGE-B) -> 42.0 (T5-B), 40.8 (BERT-B) | Hurt |
| use the crafted Aligned description prompt instead of Naive prompt (ablation) | ProLab with GPT-3.5 · ADE20K | mIoU: 47.5 -> 47.7 | Improved |
| use GPT-3.5 instead of LLAMA2-7B for descriptions (ablation) | ProLab · ADE20K | mIoU: 47.0 -> 47.5 | Improved |
| apply sigmoid with cosine similarity loss instead of BCE (ablation) | ProLab · ADE20K | mIoU: 47.4 -> 47.7 | Improved |
| cosine similarity loss without sigmoid vs BCE (ablation) | ProLab · ADE20K | mIoU: 47.4 -> 47.3 | No clear effect |
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.
- SegFormer — SegFormer: A Topic Segmentation Model with Controllable Range of Attention (AAAI 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
- Scaling Up Multi-domain Semantic Segmentation with Sentence Embeddings (IJCV 2024 · similarity 0.89)
- Language-driven Semantic Segmentation (ICLR 2022 · similarity 0.89)
- Concept Mask: Large-Scale Segmentation from Semantic Concepts (ECCV 2018 · similarity 0.89)
- Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training (AAAI 2024 · similarity 0.88)
- Training-free Open-Vocabulary Semantic Segmentation via Diverse Prototype Construction and Sub-region Matching (AAAI 2025 · similarity 0.87)
- LPOSS: Label Propagation Over Patches and Pixels for Open-vocabulary Semantic Segmentation (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.
