The predictions are different from darknet YOLOv4

[thotsawut]

I have trained YOLOv4 from the darknet of AlexeyAB repository with my custom dataset

after that i used the save_model.py to convert weight to tensorflow but the prediction is different from the prediction from darknet

some objects have lower confident, some objects have no bounding box

I think there are some parameter between these two that are different

Is there anyone facing the same problem ?

[iFlyyyy]

same problem

[ItsMeTheBee]

same here but I´m using a tiny yolo v3 model.... confidence is a lot lower and the bounding boxes often don´t really fit... the center seems to be in the right place but height and width are off

[vqbang]

same problem with yolo v3 full model, hope some quick fixes on going :( @hunglc007

[yingtina]

same problem

[thotsawut]

As I have searched for the answer for a while. I think it's because of the transformation process from the weight file to tensorflow model is not perfect originally from the library. because I have found some of the these problem in other model too. And I have found two method to relieve this problem. First is to transfrom the weight in to tensorflow model into multiple size of input for example 480x480 and 960x960 and then combine the result before feeding to nms function. this is becuase I have found that the vary input size give the different result. Second is to use the weight file directly with opencv library but this method give the worse performance in term of time but the result is the same when we test the image with original YOLO

[zz100chan]

Second is to use the weight file directly with opencv library but this method give the worse performance in term of time but the result is the same when we test the image with original YOLO

can you further explain on this method and how to implement it? Thank you so much!