Open JesseYang opened 7 years ago
Thank you for pointing out this. I read the darknet code again and think that you are right. The darknet uses anchors to match gt_boxes and pred_boxes (https://github.com/pjreddie/darknet/blob/master/src/region_layer.c#L264) but uses ious between gt_boxes and pred_boxes as targets (https://github.com/pjreddie/darknet/blob/master/src/region_layer.c#L280).
I will fix this and try to train the network later.
@longcw I am curious if the bug has been fixed. Thanks!
Any updated on this?
@longcw is the bug fixed? And what mAP do you get now?
The bug has been fixed very early at commit edda7d48ecf67954414923bc94f05ea036e1a9c3
In the code, the ground truth value for iou prediction is the iou between the anchor and the ground truth box (https://github.com/cory8249/yolo2-pytorch/blob/master/darknet.py#L111). But the paper says that " the objectness prediction still predicts the IOU of the ground truth and the proposed box ...". In my understanding the "proposed box" here is the predicted box, not the anchor box.