search

Sense-X / Co-DETR

[ICCV 2023] DETRs with Collaborative Hybrid Assignments Training
MIT License
950 stars 100 forks source link

Please help: Co-Detr can't work with single class object detection! #101

Open foresee-ai opened 8 months ago

foresee-ai commented 8 months ago

Hi,

Thank you very much for the excellent work. I have tried to train Co-Detr with customized single-class dataset (in COCO's annotation format). Although it works, the evaluation mAP is always low. I stopped the training after 8 epoch training where mAP only reached 0.01, i.e., 1%. From my understanding, original DETR's architecture define 92 classes for COCO dataset because the authors want to include the "background" object. So my question is (1) Does Co-DETR uses DETR's idea of taking the "background" object into consideration? (2) Is there a way to make Co-DETR work on single-calss object detections? Thank you!

Heidi Hu

TempleX98 commented 8 months ago

(1) No. (2) The current implementation sets num_classes=80. You can use num_classes=1 to perform single-class detection.