Research map / Detection and segmentation / Open-set and few-shot detection
Open Vocabulary Scene Parsing
Hang Zhao, Xavier Puig, Bolei Zhou, Sanja Fidler, Antonio Torralba
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 |
|---|---|---|---|
| add hypernym scoring in place of L2 scoring (joint training) (ablation) | Joint model · ADE20K zero-shot (500 infrequent classes) | hierarchical F-score: 0.5655 -> 0.6174 | Improved |
| add hypernym scoring in place of L2 scoring (image stream only) (ablation) | Image-stream model · ADE20K zero-shot (500 infrequent classes) | hierarchical F-score: 0.5375 -> 0.5937 | Improved |
| train concept stream jointly with image stream instead of fixed concept embeddings (hypernym score) (ablation) | Hyper model · ADE20K zero-shot (500 infrequent classes) | hierarchical F-score: 0.5937 -> 0.6174 | Improved |
| train concept stream jointly with image stream (L2 score) (ablation) | L2 model · ADE20K zero-shot (500 infrequent classes) | hierarchical F-score: 0.5375 -> 0.5655 | Improved |
| joint hypernym model instead of Softmax baseline | Joint-Hyper · ADE20K zero-shot (500 infrequent classes) | information content ratio: 0.1632 -> 0.2226 | Improved |
| joint training instead of image-only training (cosine score) (ablation) | Cosine model · ADE20K SceneParse150 (150 training classes) | mean IoU: 0.2803 -> 0.3152 | Improved |
| joint training instead of image-only training (L2 score) (ablation) | L2 model · ADE20K SceneParse150 (150 training classes) | mean IoU: 0.2174 -> 0.2692 | Improved |
| use cosine instead of L2 scoring in joint model (ablation) | Joint model · ADE20K SceneParse150 (150 training classes) | mean IoU: 0.2692 -> 0.3152 | Improved |
| softmax loss and domain-finetuned word vectors in Word2Vec+ vs Word2Vec (ablation) | Word2Vec regression · ADE20K SceneParse150 (150 training classes) | mean IoU: 0.2918 -> 0.3160 | Improved |
| open-vocabulary joint model vs Softmax on closed 150-class parsing | Joint-Cosine · ADE20K SceneParse150 (150 training classes) | pixel accuracy: 73.55% -> 73.15% | Hurt |
| use Word2Vec+ vs Word2Vec on zero-shot (ablation) | Word2Vec regression · ADE20K zero-shot (500 infrequent classes) | hierarchical F-score: 0.5507 -> 0.5671 | Improved |
| increase number of training classes beyond 500 (ablation) | Joint-Hyper · ADE20K zero-shot (500 infrequent classes) | hierarchical F-score | No clear effect |
Closest papers by meaning
- Towards Open Vocabulary Learning: A Survey (TPAMI 2024 · similarity 0.89)
- PLA: Language-Driven Open-Vocabulary 3D Scene Understanding (CVPR 2023 · similarity 0.88)
- On Vocabulary Reliance in Scene Text Recognition (CVPR 2020 · similarity 0.87)
- A Simple Baseline for Open-Vocabulary Semantic Segmentation with Pre-trained Vision-Language Model (ECCV 2022 · similarity 0.87)
- OpenWorldSAM: Extending SAM2 for Universal Image Segmentation with Language Prompts (NeurIPS 2025 · similarity 0.87)
- Towards Open-Vocabulary Semantic Segmentation Without Semantic Labels (NeurIPS 2024 · similarity 0.86)
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.
