Why doesn't MRCNN need to convert the operation of CropAndResize?

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Hi all,

I have a question about why MRCNN doesn’t need to convert the operation of CropAndResize in ~/samples/sampleUffMaskRCNN/converted/config.py.

Because I saw the sample of FasterRCNN which substitutes the CropAndResize operation for original operation in ~/samples/sampleUffFasterRCNN/config.py.

We can also find the crop_and_resize function in Model.py of matterport/MRCNN repo, so it should be replaced this operation, no?

Is there any reason here?

Thank you so much.

2

Hi,

We also replaced “CropAndResize” in MRCNN sample. We replace it as a whole, it is in PyramidROIAlign.

https://github.com/NVIDIA/TensorRT/blob/master/samples/opensource/sampleUffMaskRCNN/converted/config.py#L37
https://github.com/NVIDIA/TensorRT/blob/master/samples/opensource/sampleUffMaskRCNN/converted/config.py#L43

“roi_align_classifier” and “roi_align_mask” contains CropAndResize

Thanks