nms() in do_detect() in utils.py should perform per class per image, no?

[love2learn-7179]

I thought we should perform nms per class per image, no?
def nms(boxes, nms_thresh): in tool/utils.py

[HotaekHan]

ordinally NMS perform as class agnostic because of overlapped boxes.

[liuxufenfeiya]

in author'nms, too many 'for' caused the speed too slow
my P4000 GPU result:
darknet: 30 fps
tf-yolov4: 25 fps
pytorch-yolov4: 12 fps

[Tianxiaomo]

Thank you very much for your suggestions. I will optimize the NMS in the future, and you are welcome to join us.
@love2learn-7179 @HotaekHan @liuxufenfeiya

[GlassyWing]

it needs performs for each class, but this can be optimized by adding offset, For details, please refer to the nms implementation in torchvision

[HotaekHan]

@GlassyWing Could you point out the code?

[GlassyWing]

@GlassyWing Could you point out the code?

The code could be see at
https://github.com/pytorch/vision/blob/3d65fc6723f1e0709916f24d819d6e17a925b394/torchvision/ops/boxes.py#L45

The comments have explained the process in more detail