Research map / Detection and segmentation / Open-set and few-shot detection
WeDetect: Fast Open-Vocabulary Object Detection as Retrieval
Shenghao Fu, Yukun Su, Fengyun Rao, Jing LYU, Xiaohua Xie, Wei-Shi Zheng
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 coarse-grained labels to multi-granularity training (ablation) | WeDetect-Base · LVIS minival | AP: 46.4 -> 47.3 | Improved |
| add fine-grained labels to multi-granularity training (ablation) | WeDetect-Base · LVIS minival | AP: 45.1 -> 47.3 | Improved |
| use staged-wise training (initialize neck/head first) (ablation) | WeDetect-Base · LVIS minival | AP: 45.5 -> 47.3 | Improved |
| use sigmoid focal loss instead of BCE (ablation) | WeDetect-Ref 4B · COCO | AP: 43.1 -> 47.5 | Improved |
| add negative detection data (ablation) | WeDetect-Ref 4B · COCO | AP: 42.1 -> 47.5 | Improved |
| increase tokens per object from 1 to 25 (ablation) | WeDetect-Ref 4B · COCO | AP: 47.5 -> 47.6 | No clear effect |
| shuffle proposal order (ablation) | WeDetect-Ref 4B · COCO | AP: 47.5 -> 47.6 | No clear effect |
| replace fusion-based detectors with non-fusion dual-tower WeDetect-Tiny | WeDetect-Tiny vs YOLO-World-L · LVIS minival | AP: 35.4 -> 37.4 | Improved |
| use LLM as one-pass classifier over proposals instead of next-token prediction | WeDetect-Ref 4B vs Qwen3-VL 4B · RefCOCO/+/g average | accuracy: 86.6 -> 93.2 | Improved |
| use object-level cached embeddings (WeDetect-Uni) vs image-level CLIP-style retrieval | WeDetect-Large-Uni vs FG-CLIP2 · COCO object retrieval | F1: 57.7 -> 83.6 | Improved |
| use WeDetect-Large-Uni vs UPN on LVIS at 100 proposals | WeDetect-Large-Uni · LVIS | AR50: 65.0 -> 70.9 | Improved |
| use WeDetect-Large-Uni vs UPN on COCO AR50 at 100 proposals | WeDetect-Large-Uni · COCO | AR50: 90.6 -> 89.7 | Hurt |
Closest papers by meaning
- SHiNe: Semantic Hierarchy Nexus for Open-vocabulary Object Detection (CVPR 2024 · similarity 0.88)
- Learning Background Prompts to Discover Implicit Knowledge for Open Vocabulary Object Detection (CVPR 2024 · similarity 0.87)
- DitHub: A Modular Framework for Incremental Open-Vocabulary Object Detection (NeurIPS 2025 · similarity 0.87)
- The Devil is in the Fine-Grained Details: Evaluating Open-Vocabulary Object Detectors for Fine-Grained Understanding (CVPR 2024 · similarity 0.87)
- GUIDED: Granular Understanding via Identification, Detection, and Discrimination for Fine-Grained Open-Vocabulary Object Detection (NeurIPS 2025 · similarity 0.87)
- Language-conditioned Detection Transformer (CVPR 2024 · 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.
