search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.63k stars 7.95k forks source link

Image Selection For YOLOv4 and YOLOv4-Tiny #7755

Open varungupta31 opened 3 years ago

varungupta31 commented 3 years ago

I'm trying to train custom YOLOv4 and YOLOv4-Tiny models for a small object (Human nose). I have images of it in various orientations but at primarily two distances from the camera - closer (around 1600 images) and further (around 2600 images).

To make the model accurate enough to handle detection at various distances (or in general, a good nose detector model based on the above data I have), How shall I distribute my training set for optimal results? I'll be having only the class 'human nose' in the detector, so will the split matter?

Also, I read that the suggested volume is 2000 images per class, would using around 4200 lead to over-fitting?

Do these facts apply similarly for an optimal YOLOv4-Tiny model, or some manipulations may be needed for that ? (e.g. increasing the training data size for tiny, etc.)

Thanks.

stephanecharette commented 3 years ago

This seems to be a duplicate of https://github.com/AlexeyAB/darknet/discussions/7754.