Skip to content

Latest commit

 

History

History
32 lines (24 loc) · 3.3 KB

README.md

File metadata and controls

32 lines (24 loc) · 3.3 KB
Raw

Distilling DETR with Visual-Linguistic Knowledge for Open-Vocabulary Object Detection

Results and Models

LVIS

We train the model on LVIS dataset with only base-category annotations, and validate the model on LVIS v1 val with both base and novel categories. The text prompts, provided by DetPro, used for LVIS dataset is same as in ViLD.

Model mask APr / APc / APf / AP bbox APr / APc / APf / AP Config Text Prompt Download
DK-DETR 20.5 / 29.0 / 35.3 / 30.0 22.4 / 31.9 / 40.1 / 33.5 config Google Drive Google Drive | BaiduYun

Generalization Ability

To demonstrate the generalization ability of the open-vocabulary object detection model, we directly evaluate the LVIS-trained model on COCO, Objects365 and Pascal VOC datasets.

Model Dataset AP AP50 AP75 Config Text Prompt Download
DK-DETR COCO 39.3 54.5 42.8 config Google Drive Google Drive | BaiduYun
DK-DETR Objects365 13.0 17.9 13.9 config Google Drive Google Drive | BaiduYun
DK-DETR Pascal VOC - 71.1 61.3 config Google Drive Google Drive | BaiduYun

Citation

@inproceedings{li2023distilling,
  title={Distilling DETR with Visual-Linguistic Knowledge for Open-Vocabulary Object Detection},
  author={Li, Liangqi and Miao, Jiaxu and Shi, Dahu and Tan, Wenming and Ren, Ye and Yang, Yi and Pu, Shiliang},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={6501--6510},
  year={2023}
}