What are the differences between the batched_nms calculation logic of torchvision 0.8.0 and the calculation logic of the nonmaxsuppression operator of ONNX?

1

Hi, I find that when I run torchvision’s nms and onnx’s nms single operator using the same input, the output of the two operators is inconsistent, and the output of onnx’s nms is a subset of the output of torchvision’s nms. Can someone tell us the difference between the logic of calculation? In my knowledge, the calculation logic of nms is consistent, but the inconsistency of results is puzzling.