Closed silicon2006 closed 3 years ago
@silicon2006 we output object confidence times class confidence times predicted IoU.
I think yolov3 and yolov4 are outputing something a bit simpler (and less performant).
iou aware. And score = sqrt(pred_iou*cls) can get score with 0.9+++
Good Idea, Now the socre (object confidence times class probability) can reach 0.95+ when testing on the train dataset; Thanks!
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.
@Aktcob what's pred_iou when inference?
@jinfagang if training is too short, predicted iou will be lower, since iou is lower. This will result in lower confidences as well, so low confidences during inference is a good indicator that you have not trained long enough.
I am using default pretrained resnet101 and scores are still too small.
@silicon2006 we output object confidence times class confidence times predicted IoU.
I think yolov3 and yolov4 are outputing something a bit simpler (and less performant).
So how can I cancel this setting? Output objectness confidence score directly without scaled by IOU and class confidence
@Steinwang objectness targets can be modified in the loss function: https://github.com/ultralytics/yolov5/blob/71621df87589faea19ba4c4098bb68e73201f30c/utils/loss.py#L144-L152
❔Question
When testing, the confidence or category probability value is too small, even when testing on the training data set, the confidence value is also small. However, the confidence and category probability of the darknet version of yolov4 and yolov3 can reach 0.99 (on the training data set), why? and how to fix it?
Additional context