BatchedNms plugin issue

[niaoyu]

Hi, in the trt document, The batchedNMSPlugin should take two inputs, boxes input and scores input:

1. Boxes , shape [batch_size, number_boxes, number_classes, number_box_parameters]
2. Scores, shape [batch_size, number_boxes, number_classes]

Here in the repo, it is right for scores https://github.com/TrojanXu/yolov5-tensorrt/blob/bab1ff1b5fcf7b8499eb8df1560d50ee205ee891/main.py#L199

But for boxes, why the number_classes=1? https://github.com/TrojanXu/yolov5-tensorrt/blob/bab1ff1b5fcf7b8499eb8df1560d50ee205ee891/main.py#L206

Furthermore, if my class is large than 1- example coco dataset, do i need to broadcast the Boxes dimenson from 1 to 80?

[niaoyu]

Oh, I see you just use the param shareLocation=1, So the classnum should be 1. Am I right?

What‘s more, is it will be OK if I transfer most part of post trt processing to torch, and only keep the trt BatchedNmsPlugin here to make the code simple？

[TrojanXu]

Yes, the bbox can be shared across different classes. Sure. What's more, I think even the TRT batchedNMS can be dropped if you choose to use torch implementation so that you can use fast nms or other nms to meet your requirements.

[niaoyu]

Thanks for your reply. We have the demand that process the yolov5 model both in python and c++. So it will be suitable if wo could process them the same way.

I figure out the transformation yesterday, the performance here： (v100，trt7，yolov5X，batchsize=1，reso640*384，without resizing) FPS torch: 38.62693323274236 FPS tensort: 130.47280364924086