Research map / Detection and segmentation / Open-set and few-shot detection
Going Denser with Open-Vocabulary Part Segmentation
Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, Zhicheng Yan
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 parsed part annotations of novel objects (Parsed IN-S11) to training (ablation) | VLPart trained on Pascal Part · PartImageNet val, 40 parts (cross-dataset) | mAP: 4.5 -> 7.8 | Improved |
| add image-level IN-S11 labels without parsing (ablation) | VLPart trained on Pascal Part · PartImageNet val, 40 parts | mAP: 4.5 -> 5.4 | Improved |
| add LVIS and PACO part-level data (ablation) | VLPart trained on Pascal Part · PartImageNet val, 40 parts | mAP: 4.5 -> 7.8 | Improved |
| add VOC object-level data (ablation) | Pascal Part base-part model · Pascal Part novel parts | AP50: 3.7 -> 5.9 | Improved |
| add image-level IN-S20 labels on top of object data (ablation) | VLPart on Pascal Part base split · Pascal Part novel parts | AP50: 5.9 -> 3.1 (novel); 36.8 -> 37.5 (all) | No clear effect |
| add parsed IN-S20 part annotations (ablation) | VLPart on Pascal Part base split · Pascal Part novel parts | AP50: 3.1 -> 11.0 | Improved |
| joint training on LVIS, PACO, Pascal Part, PartImageNet (no extra image data) (ablation) | VLPart ResNet50 · PACO and PartImageNet vs dataset-specific | AP: PACO 9.3 vs 10.6; Pascal Part 22.6 vs 19.4 | Mixed |
| add IN and Parsed IN to joint training (ablation) | VLPart ResNet50 joint · PartImageNet, Pascal Part, PACO | AP: PartImageNet 29.1 -> 31.6; Pascal Part 22.6 -> 24.0; PACO 9.3 -> 9.6 | Improved |
| scale backbone from ResNet50 to Swin-B Cascade Mask R-CNN | joint VLPart · PACO | AP: 9.6 -> 15.6 | Improved |
| use max-size alignment of image labels instead of parsing (ablation) | VLPart cross-dataset · PartImageNet, 40 parts | mAP: 5.4 -> 5.3 | Hurt |
| use max-score alignment instead of parsing (ablation) | VLPart cross-dataset · PartImageNet, 40 parts | mAP: 5.4 -> 6.0 (parsing: 7.8) | Improved |
| swap prompt template to '[part] of a [object]' (ablation) | fully supervised part detector · Pascal Part vs PartImageNet | AP: Pascal 19.1 -> 18.4; PartImageNet 29.7 -> 29.9 | Mixed |
Closest papers by meaning
- Fine-Grained Image-Text Correspondence with Cost Aggregation for Open-Vocabulary Part Segmentation (CVPR 2025 · similarity 0.92)
- LangHOPS: Language Grounded Hierarchical Open-Vocabulary Part Segmentation (NeurIPS 2025 · similarity 0.92)
- Understanding Multi-Granularity for Open-Vocabulary Part Segmentation (NeurIPS 2024 · similarity 0.91)
- Training-free Open-Vocabulary Semantic Segmentation via Diverse Prototype Construction and Sub-region Matching (AAAI 2025 · similarity 0.90)
- The Devil is in the Fine-Grained Details: Evaluating Open-Vocabulary Object Detectors for Fine-Grained Understanding (CVPR 2024 · similarity 0.89)
- OpenWorldSAM: Extending SAM2 for Universal Image Segmentation with Language Prompts (NeurIPS 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.
