ops.nms get wrong results when set iou_threshold=0

[ccblublu]

🐛 Describe the bug

boxes = torch.tensor([[       2748,         618,        2830,         659],
       [       2690,         642,        2749,         681],
       [       2690,         639,        2746,         681]])

score = torch.tensor([1,1,0.5])

iou =  torchvision.ops.box_iou(boxes, boxes)

index1 =  torchvision.ops.nms(boxes, socre, iou_threshold=0)

index2 = torchvision.ops.nms(boxes, socre, iou_threshold=0.1)

(index1==index2).all()

when set iou_threshold=0, then index return is wrong:
i set boxes
([[ 2748, 618, 2830, 659],
[ 2690, 642, 2749, 681],
[ 2690, 639, 2746, 681]])
from ops.box_iou:
([[1.0000, 0.0030, 0.0000],
[0.0030, 1.0000, 0.8846],
[0.0000, 0.8846, 1.0000]] is normal
but when run nms with score [1,1,0.5]
get result:
tensor([0, 2])
while set iou_threshold=0.1,it return:
tensor([0, 1]) is correct

Versions

torch==2.4.1
torchvision==0.19.1
CUDA==12.1