search

facebookresearch / detr

End-to-End Object Detection with Transformers
Apache License 2.0
13.59k stars 2.45k forks source link

Got low confidence scores for single class detection #522

Open yiyingsheng07 opened 2 years ago

yiyingsheng07 commented 2 years ago

Hi, I used DETR for single class detection on my own dataset, but I got some problems when doing predictions. All scores for the detected object from 20 queries are quite low (~-13) while the "no-object" class has the hightest score (~4). So after Softmax, the confidence score for the object comes to ~4e-08. However, the location of bboxes looks right. I was wondering why I got low confidence scores? Should I try to change <eos_coef>?

Thanks in advance!

AlessandroMagliola commented 2 years ago

I have the same problem. I built my own dataset and implemented a single class object detection, trained 80 epochs and scores are very very low. Moreover, I am using the following version: torch=1.12.0, transformers=4.20.1, cuda= 11.1