[Fix] batched_nms using proper box score values

[ilijad]

In the bbox_nms the box classes valid_cls were provided to batched_nms instead of box scores valid_con

[tloki]

can you put named arguments for the sake of readability?

Suggested change

	nms_idx = batched_nms(valid_box, valid_con, batch_idx, nms_cfg.min_iou)
	nms_idx = batched_nms(boxes=valid_box, scores=valid_con, idxs=batch_idx, iou_threshold=nms_cfg.min_iou)

[ilijad]

done

[ilijad]

Problem example

import torch
from torchvision.ops import batched_nms # type: ignore

valid_box = torch.Tensor([[10,10,20,20],[12,12,22,21]])
valid_con =torch.Tensor([0.3, 0.9])
batch_idx = torch.Tensor([0,0])
valid_cls = torch.Tensor([1,1])

rez_con = batched_nms(valid_box,valid_con,batch_idx, 0.5)
rez_cls = batched_nms(valid_box,valid_cls,batch_idx, 0.5)

print(rez_con, rez_cls) # rez_con = 1 rez_cls = 0
valid_con_reversed = torch.Tensor([0.9,0.3])

rez_con = batched_nms(valid_box,valid_con_reversed,batch_idx, 0.5)
rez_cls = batched_nms(valid_box,valid_cls,batch_idx, 0.5)
print(rez_con, rez_cls) # rez_con = 0 rez_cls = 0

different results obtained because of valid_box box order

[henrytsui000]

Hi,

I believe this issue has already been fixed in PR #76. However, the DEPLOY branch hasn't been merged into the main branch yet. It will be merged later.

Best regards,
Henry Tsui

[ilijad]

Yes you are right. I'll close the pull request.