What's the post process difference between V3 and V4

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )

http://pjreddie.com/darknet/

Other

21.65k stars 7.96k forks source link

What's the post process difference between V3 and V4 #5784

Open Amore-HDU opened 4 years ago

Amore-HDU commented 4 years ago

@AlexeyAB Hello， First, thans to your greate work!!! I have trained YoloV4 in my custom dataset. In convenient of deploying, I have changed mish to relu and set scale_x_y=1. I tried to convert yolov4 to caffe as same as yolov3. It converted successfully,but when I use yolov3 post process to run the converted v4 model,I found the precision fall 5%. So I want to know What's the post process difference between V3 and V4

VolkovAK commented 4 years ago

@Amore-HDU , did you retrain model after these changes? IMHO wrong post-processing should drop precision much more than 5%, but i'm not sure. Can you show an example of detection?

Amore-HDU commented 4 years ago

@VolkovAK

changed mish to relu and retrain, the precision nearly remain unchaned
train with scale_x_y=1.2 1.1 1.05, test with scale_x_y=1.0 1.0 1.0, I found the precision nearly remain unchanged Due to second phenomenon, I think yolov4's post process is the same as yoloV3.So I use yolov3's post process.

AlexeyAB commented 4 years ago

How did you check precision?
What mAP did you get after these changes?
did you retrain model after these changes?

Amore-HDU commented 4 years ago

@AlexeyAB

How did you check precision?
- I have converted weight to caffe model, and run evaluation code in my owner test dataste.
What mAP did you get after these changes?
did you retrain model after these changes?
- changed mish to relu and retrain, the precision nearly remain unchaned
- train with scale_x_y=1.2 1.1 1.05, test with scale_x_y=1.0 1.0 1.0, I found the precision nearly remain unchanged