MULTI-CLASS OUTPUT

[glenn-jocher]

🚀 Feature

Multi-class output is now enabled in this repo. This issue serves to track before and after mAPs on COOC using this feature along with different NMS techniques #679.

yolov3/utils/utils.py

Line 462 in 9309d35

def non_max_suppression(prediction, conf_thres=0.5, nms_thres=0.5, multi_cls=True, method='vision'):

The test code is below, run on P100 colab instance. Results will be updated as they become available.

python3 test.py --weights ultralytics68.pt --nms-thr 0.6 --img-size 608 --device 0

	Speed mm:ss	COCO mAP @0.5...0.95	COCO mAP @0.5
'vision' nms	5:27	40.3	60.4
'vision' nms + multi_cls	10:47	40.9	61.5
'merge' nms	18:41	40.7	60.7
'merge' nms + multi_cls	15:15	41.4	61.7

[glenn-jocher]

Same at 0.55 and 0.70 for max mAP.

python3 test.py --weights ultralytics68.pt --conf 0.001 --data coco2014.data --img-size 608 --nms 0.70

	Speed mm:ss	COCO mAP @0.5...0.95	COCO mAP @0.5
'vision' nms 0.55 + multi_cls	8:25	40.8	61.6
'merge' nms 0.55 + multi_cls	21:59	41.2	61.9
'vision' nms 0.70 + multi_cls	7:58	41.1	60.9
'merge' nms 0.70 + multi_cls	19:26	41.6	61.1

[github-actions]

This issue is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in 5 days