YOLOv4 no detections - Githubissues

AlexBlck commented 4 years ago

Hi @AlexeyAB!

I have a dataset of images with lots (~1500) of small objects on them and I was previously able to train YOLOv3 to perform on it quite well.

I am now trying to use YOLOv4 with the same dataset and, even thought the loss is dropping, when I test it out, there are no detections at all, even at low thresholds.

It might very well be the case of not training long enough, but maybe you can spot a problem in my cfg file. I'm attaching both my working v3 setup and the new v4.

custom_v3_v4.zip

AlexeyAB commented 4 years ago

Show chart with Loss and mAP

AlexBlck commented 4 years ago

Here's the end part of the chart. chart After training some more I managed to get the detections, but in order to get a good recall I still have to lower the threshold to about 0.05:

Are there any ways to improve recall? More negative examples perhaps.

AlexeyAB commented 4 years ago

Just train 12000 iterations or more instead of 6000
How many images do you have?
To improve recall you need to use more positive examples.
by the way you get good accuracy 87.45% AP50

AlexeyAB / darknet

YOLOv4 no detections #5389