Dear Author,
Excellent source code you provide.
I use your code and test the yolov3.coco pre-trained model perfectly.
However, I use the same procedure on yolov3.openimages pre-trained model, the outputs are very poor: it predict so many classes and so many boxes, but none of them is correct.
And I didn't find the bug yet.
Do you know any difference between the yolov3.coco and the yolov3.openimage?
I just found that their output layers and the channel nums of one convolutional layer are different, i.e. 80vs601, 255vs1818
The pre-trained weights can be load, the networked can run, but I just didn't found the bug that causes the very poor performance.
Dear Author,
Excellent source code you provide.
I use your code and test the yolov3.coco pre-trained model perfectly.
However, I use the same procedure on yolov3.openimages pre-trained model, the outputs are very poor: it predict so many classes and so many boxes, but none of them is correct.
And I didn't find the bug yet.
Do you know any difference between the yolov3.coco and the yolov3.openimage?
I just found that their output layers and the channel nums of one convolutional layer are different, i.e. 80vs601, 255vs1818
The pre-trained weights can be load, the networked can run, but I just didn't found the bug that causes the very poor performance.