Matterport's Mask RCNN to .etlt conversion?

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

I have some mrcnn models trianed with the well known repo: GitHub - matterport/Mask_RCNN: Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow and I wanted to deploy it in a Jetson device with deepstream.

I have found here IndexError: list index out of range(Object Detection And Instance Segmentations With A TensorFlow Ma… - Jetson & Embedded Systems / Jetson AGX Xavier - NVIDIA Developer Forums that some time ago we had to convert it to .uff but now this seems deprecated.

I have been able to run succesfully a .etlt maskrcnn model with deepstream. So the questions are:

  • Should I transform my .h5 model file to .etlt? If so, how?
  • Is now .etlt files the unique way to run mrcnn models with deepstream?
  • Should I convert my .h5 to -uff despite its deprecation?

Thanks!

3

@Morganh Can you provide some suggestion?

4
  1. Currently it is not supported for mask_rcnn.
    BTW, for latest 22.05 TAO, it can support transforming classification or Unet onnx model to .etlt model. This is mentioned in BYOM section.BYOM Converter — TAO Toolkit 3.22.05 documentation
  2. No. Actually end users can generate tensort engine and deploy it directly in deepstream.
  3. Yes. You can refer to TensorRT/README.md at main · NVIDIA/TensorRT · GitHub