search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.59k stars 7.95k forks source link

Questions on the changes of Yolov4 from Yolov3 in terms of detection, bounding box, and loss functions. #6276

Open francismontalbo opened 3 years ago

francismontalbo commented 3 years ago

I would like to ask if Yolov4 still uses the previous detection approach stated in the original YOLO paper?

https://arxiv.org/pdf/1506.02640.pdf Specifically, on the Unified Detection.

With YOLOv4 implementation on this repository, thus it uses both the SPP and PAN as necks?

For the loss function, does YOLOv4 use IOU instead of the previous MSE? or are they used differently?

These questions are for further validation according to what I've read. Thank you to everyone who will answer and help. This could also be a good thread for discussions about the improvements from v3 to v4.

I' am still learning and highly interested with detection models like YOLO.

AlexeyAB commented 3 years ago

With YOLOv4 implementation on this repository, thus it uses both the SPP and PAN as necks?

Yes.

For the loss function, does YOLOv4 use IOU instead of the previous MSE? or are they used differently?

No, it uses CIoU loss for bbox regression, page 7: https://arxiv.org/pdf/2004.10934.pdf

Read:

francismontalbo commented 3 years ago

With YOLOv4 implementation on this repository, thus it uses both the SPP and PAN as necks?

Yes.

For the loss function, does YOLOv4 use IOU instead of the previous MSE? or are they used differently?

No, it uses CIoU loss for bbox regression, page 7: https://arxiv.org/pdf/2004.10934.pdf

Read:

Wonderful!Thank you for such quick response! I've seen that it still uses the same bounding box method or unified detection method with a different loss. Is that correct? Thank you, again! 😄