search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.74k stars 7.96k forks source link

Detection results #3782

Open sctrueew opened 5 years ago

sctrueew commented 5 years ago

Hi everyone,

I have 100 classesand I have ~2k images for each class. For example I get a good result when I train for classes but when I want to train for all classes some classes don't work well, while I have got a good result for those classes when I've trained for 10 classes. My iteration number for 10 calss is over 20k and for 200 class is over 200k and I'm using Yolov3 model.

Where is my problem?

Thanks in advance

dasmehdix commented 5 years ago

-Check the paths of the files -Check the label files are own the object coordinates (txt.files) -Maybe overfitting occurs on your model. -As a solution, use resized,flipped,distorted photos on some classes.

These are just some things that come my mind.

sctrueew commented 5 years ago

@dasmehdix Hi,

I'm sure of the paths and the coordinates and I'm using different images and always I set random = 1. I think, overfitting occurs on my model. How can I avoid the overfitting?

dasmehdix commented 5 years ago

@zpmmehrdad In your dataset, there can be batch of images that has very similiar images.If this situation occurs,you have to shuffle them.This is just a solution.Please, look for this link.https://towardsdatascience.com/deep-learning-3-more-on-cnns-handling-overfitting-2bd5d99abe5d