search

pjreddie / darknet

Convolutional Neural Networks
http://pjreddie.com/darknet/
Other
25.85k stars 21.33k forks source link

Yolo has 100 percent of accuracy running test command but do not detect the objects #2586

Closed Vxtzq closed 1 year ago

Vxtzq commented 1 year ago

Hello, I created a simple yoloV4 tiny with transfer learning to recognize flowers, I used 260 images with 1 class and I trained the model with following command : ./darknet detector train data/obj.data data/yolov4-tiny.cfg data/yolov4-tiny.conv.29 -dont_show -map. I trained for about 500 epochs and loss is 1. I can't upload an image because i compiled darknet with opencv but you can see micro bounding boxes with all 100 percent chance on the image, (the image is from train set, so I expect it to be good) but no. I really don't know what to do I really need help. PS : I don't have -nan values in training and it seems to be normal.

Vxtzq commented 1 year ago

In clear : I want to know if it's possible to transfer learning with yolov4-tiny and with 1 class

Vxtzq commented 1 year ago

Posted this issue on "official" darknet page https://github.com/AlexeyAB/darknet/issues/8728