search

GiovanniPasq / DA-Faster-RCNN

Detectron2 implementation of DA-Faster R-CNN, Domain Adaptive Faster R-CNN for Object Detection in the Wild
MIT License
53 stars 10 forks source link

NAN on the evaluation #19

Closed arpanpoudel closed 1 month ago

arpanpoudel commented 1 month ago

Hi, Thank you for the awesome documentation.

However, I have one issue. I am getting NAN during evaluation during training. Could you provide insight ( I am doing a single object detection)

image
GiovanniPasq commented 1 month ago

Your model has a mAP of 0.075 for the 'Broiler' class. It seems there may be issues with the annotations in your training set. Additionally, please note that a test set of only 10 images is quite small and may not provide reliable results. NaN results are due to values close to zero, whose division gives NaN when the metric is computed.

arpanpoudel commented 1 month ago

Your model has a mAP of 0.075 for the 'Broiler' class. It seems there may be issues with the annotations in your training set. Additionally, please note that a test set of only 10 images is quite small and may not provide reliable results. NaN results are due to values close to zero, whose division gives NaN when the metric is computed.

Thank you for insights. I found that the bounding box size in the test set doesn't exist for mAPs(32x32) and mAP(forgot size) and because of that, I am getting the NaN error.