search

BobLiu20 / YOLOv3_PyTorch

Full implementation of YOLOv3 in PyTorch
624 stars 191 forks source link

Hi,I have some problem to trained voc2007 dataset!!! #44

Open BCWang93 opened 5 years ago

BCWang93 commented 5 years ago

When I use the code to train the voc2007 dataset,I get the result as follows: 1 2 3 4 5 6 7 8 And the detected result is none!Anyone can help me ?Thanks!

AndrewZhuZJU commented 5 years ago

I trained the VOC2012 dataset, and everything just looks fine for me. Screenshot Did you change the class number from 80 to 20? Or you should give more details.

BCWang93 commented 5 years ago

I trained the VOC2012 dataset, and everything just looks fine for me. Screenshot Did you change the class number from 80 to 20? Or you should give more details.

Thanks for your reply! Did you use the origin code? or you have some change?I have changed the class number from 80 to 20.But I get this result in voc2007.

BCWang93 commented 5 years ago

I trained the VOC2012 dataset, and everything just looks fine for me. Screenshot Did you change the class number from 80 to 20? Or you should give more details.

Hi,can you see the loss_x and loss_y,loss_w,loss_h?it is convergence？Btw,you use the origin code? Did you changed something?

BCWang93 commented 5 years ago

I trained the VOC2012 dataset, and everything just looks fine for me. Screenshot Did you change the class number from 80 to 20? Or you should give more details.

Can you detect the bbox use the model of you trained?When I test the trained model use image,the ouput is none,it can't detected anything!Thanks!

AndrewZhuZJU commented 5 years ago

I trained the VOC2012 dataset, and everything just looks fine for me. Screenshot Did you change the class number from 80 to 20? Or you should give more details.

Can you detect the bbox use the model of you trained?When I test the trained model use image,the ouput is none,it can't detected anything!Thanks!

No need to change the code. The reason for nothing detected is NMS removes all your predicted bounding box, which means your model is not trained well.

BCWang93 commented 5 years ago

I trained the VOC2012 dataset, and everything just looks fine for me. Screenshot Did you change the class number from 80 to 20? Or you should give more details.

Can you detect the bbox use the model of you trained?When I test the trained model use image,the ouput is none,it can't detected anything!Thanks!

No need to change the code. The reason for nothing detected is NMS removes all your predicted bounding box, which means your model is not trained well.

OK，I don't change the code.but the result I get is not well like you.I don't know the reason is what? This trouble me some days! Did you use the data_agumentent.py keepAspect in this code?Thanks!

eng100200 commented 4 years ago

@BCWang93 @AndrewZhuZJU i think both of you have used VOC2007? did you have use coco_person_keypoints_train_2017?