Research map / Detection and segmentation / Open-set and few-shot detection
EdaDet: Open-Vocabulary Object Detection Using Early Dense Alignment
Cheng Shi, Sibei Yang
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 object-level alignment with dense-level alignment (ablation) | EdaDet · COCO | novel box AP50: 15.1 -> 30.1 | Improved |
| add top-k mask M_k (k=12x12) over box scores (ablation) | EdaDet dense alignment · COCO | novel box AP50: 30.1 -> 33.1 | Improved |
| use a small top-k mask (7x7) instead of 12x12 (ablation) | EdaDet dense alignment · COCO | novel box AP50: 33.1 -> 29.3 | Hurt |
| fuse CLIP dense probability map (lambda=0.25) (ablation) | EdaDet · COCO | novel box AP50: 33.1 -> 35.7 | Improved |
| add global alignment to CLIP image feature (ablation) | EdaDet · COCO | novel box AP50: 35.7 -> 37.8 | Improved |
| deeply decouple proposal generation and classification branches (ablation) | CLIP-classifier detector · COCO | novel box AP50: 7.6 -> 14.2 | Improved |
| deeply decoupled DETR-style proposal head instead of Mask-RCNN RPN or Def-DETR (ablation) | proposal generation · LVIS base categories | top-300 AR: 0.33 -> 0.36 | Improved |
| add prompt ensembling to the classifier (ablation) | decoupled CLIP detector · COCO | novel box AP50: 14.2 -> 15.1 | Improved |
| apply dense alignment only at one early stage (conv2_x or conv3_x) instead of both (ablation) | EdaDet · COCO | novel box AP50: 37.8 -> 35.9 / 36.2 | Hurt |
| add self-training (ablation) | EdaDet · COCO | novel box AP50: 37.8 -> 40.2 | Improved |
| train longer (10 to 30 epochs) with ViT-Base | EdaDet-ViT-Base · LVIS | novel mask AP: 29.9 -> 35.6 | Improved |
Later papers that compared with this method
2 later accepted papers report EdaDet as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection (ECCV 2024 · better on 2 of 2 shared results)
- InstaGen: Enhancing Object Detection by Training on Synthetic Dataset (CVPR 2024 · 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
- Bridging the Gap between Object and Image-level Representations for Open-Vocabulary Detection (NeurIPS 2022 · similarity 0.93)
- Exploring Region-Word Alignment in Built-in Detector for Open-Vocabulary Object Detection (CVPR 2024 · similarity 0.90)
- Open-Det: An Efficient Learning Framework for Open-Ended Detection (ICML 2025 · similarity 0.90)
- LLMs Meet VLMs: Boost Open Vocabulary Object Detection with Fine-grained Descriptors (ICLR 2024 · similarity 0.90)
- DitHub: A Modular Framework for Incremental Open-Vocabulary Object Detection (NeurIPS 2025 · similarity 0.90)
- CLIPSelf: Vision Transformer Distills Itself for Open-Vocabulary Dense Prediction (ICLR 2024 · 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.
