Possible bugs in YOLO Layer's forward about pytorch-yolov3 HOT 5 CLOSED

I think you are right on both points. I modified my code as you suggested but haven't tried training yet.

from pytorch-yolov3.

tconf and mask are defined same in utils

from pytorch-yolov3.

Yes we can actually remove tconf and just use (1 - mask)

from pytorch-yolov3.

Actually removing mask * (1 - mask) is a problem. You want the noobject conf target to be made of zeros since you want them to have a low confidence. The way it was is correct I think.

from pytorch-yolov3.

1. This is fixed in 959e0ff. Before this commit confidence was handled as in YOLOv2. This commit implement the confidence mask according to version 3 where the ignore threshold is used to filter out predictions where the anchor has an overlap with the target larger than 0.5.
2. This should be fixed in 1dcdf11

from pytorch-yolov3.