search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.58k stars 7.95k forks source link

Could you please give some references that use objectness smooth? #7226

Closed wtiandong closed 3 years ago

wtiandong commented 3 years ago

❔Question

Could you please give some references that use IOU as objectness labels?

Additional context

I notice it is called objectness smooth in yolov4 codes, but I can't find who is first using this method or in any papers, not even in the yolov4 paper. Thank you.

wtiandong commented 3 years ago

Refer to https://github.com/ultralytics/yolov5/issues/1863

AlexeyAB commented 3 years ago

This is from YOLOv2 - 2016 year: https://github.com/AlexeyAB/darknet/blob/b11af9e93b4ec225450b0dd73bdd7aaa1f2912b8/src/region_layer.c#L351

YOLO9000: Better, Faster, Stronger 25 Dec 2016: https://arxiv.org/pdf/1612.08242.pdf

Following YOLO, the objectness prediction still predicts the IOU of the ground truth and the proposed box and the class predictions predict the conditional probability of that class given that there is an object.

Losses: image

wtiandong commented 3 years ago

OK, Thank you! @AlexeyAB It's interesting that in Yolov3 the IOU in objectness is removed, and now back in Yolov4.

https://github.com/AlexeyAB/darknet/blob/17520296c730c7d7e2683452b11bf50fc8959688/src/yolo_layer.c#L223

AlexeyAB commented 3 years ago
wtiandong commented 3 years ago
  • objectness_smooth slightly increases AP (AP50...95) but decreases AP50
  • objectness_smooth was forcibly enabled in YOLOv2, while it is optional in YOLOv4 (enabled in Scaled-YOLOv4 and disabled in YOLOv4)

OK，thank you!