search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.57k stars 7.95k forks source link

Training yolov3 with one class #3959

Open 20162021 opened 4 years ago

20162021 commented 4 years ago

So, I need to detect pedestrians in lwir pictures. I have about 50k annotate photos with pedestrian and just with backgrounds. I have found that for only one class you need 4000-5000 iterations but it's just for 2k photos. So, do I need more iterations or 5k would be enough ?

Knuust commented 4 years ago

I would observe the loss and mAP rate. If it doesn't decrease/increase anymore, you should stop training. In general: the bigger the dataset, the more you can train on the set without having problems with overfitting. To come to a point: after 5k Iterations you will probably have an ok result. But you can probably improve it by continuing training.