COCO eval results boost by changing NMS function

[mikel-brostrom]

Hi @Tianxiaomo , I just ported your Yolov4 to Pytorch Lightning. I get the following results evaluating with pycocotools

	mAP@0.5	mAP@0.5:0.95
TianXiaomo(pycocotools evaluation)	0.704	0.471
Ported to Pytorch Lightning using Yolov5 NMS (pycocotools evaluation)	0.730	0.491
Original Yolov4 (custom evaluation)	0.710	0.466

The only thing I changed was the NMS function. I used this one from Yolov5 and adapted the head to be compatible with it. The boosts are substantial.

[PrajwalCogniac]

Hey @mikel-brostrom
Can you please let me know what changes did you make to get the inference boost ?