search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.77k stars 7.96k forks source link

Low mAP of yolo-tiny-prn. #4091

Open ChenCong7375 opened 5 years ago

ChenCong7375 commented 5 years ago

I trained yolo-tiny-prn.cfg on VOC dataset with yolo-tiny-conv.13 as pretrain model. I got 66%mAP on VOC of yolov3-tiny.cfg But yolov3-tiny-prn.cfg got a lower mAP . Have you ever trained it on your own dataset?

AlexeyAB commented 5 years ago

What mAP did you get on?

  1. yolov3-tiny.cfg
  2. yolo-tiny-prn.cfg
  3. yolov3-tiny-prn.cfg
ChenCong7375 commented 5 years ago

@AlexeyAB yolov3-tiny.cfg:320320,66%mAP yolov3-tiny-prn.cfg:320320,60%mAP

AlexeyAB commented 5 years ago

How many iterations did you train both cfg-files?

ChenCong7375 commented 5 years ago

50000 iterations，for 20 classes in VOC dataset

WongKinYiu commented 5 years ago

i think the reason is that u use the coco pretrained model on yolo-v3-tiny instead of yolo-v3-tiny-prn. i use prn.conv.15 instead of tiny.conv.13, and the results are as follows: image

ChenCong7375 commented 5 years ago

@WongKinYiu thanks very much for your repo.

ChenCong7375 commented 5 years ago

@WongKinYiu So you used prn.conv.15 and get 64% mAP? But I got 68% mAP… so can you share your chart?

ChenCong7375 commented 5 years ago

prn416 @WongKinYiu

WongKinYiu commented 5 years ago

i do not save the chart. and i trained the model in march, there are so many improvement of this repo in these days.

however, 68% mAP is better than 64%, it is a good news.

samohadid commented 3 years ago

hi @WongKinYiu can you give me a link to download prn.conv.15 pre-trained model ? thank you very much

developer0hye commented 3 years ago

@WongKinYiu

Thank you for experiments!

I have two quesetions.

  1. What does W/O COCO model mean?

training from imagenet pretrained model? or training from scratch?

  1. Normalization method for preprocessing

Did you normalize the input image by mean and std value?

If you did, what do you use mean and std from imagenet or voc dataset?

WongKinYiu commented 3 years ago

What does W/O COCO model mean?

w\o = without, without coco pretrain.

training from imagenet pretrained model? or training from scratch?

from imagenet pretrain

Normalization method for preprocessing

/255.

developer0hye commented 3 years ago

@WongKinYiu Thanks!

developer0hye commented 3 years ago

@WongKinYiu

I have new two question.

  1. How do you normalize image to train the model on Imagenet?

same with /255.?

  1. What resize method do you use for preprocessing? https://github.com/AlexeyAB/darknet/issues/232#issuecomment-336955485