Mask RCNN (mobilenet backbone) with tensorRT on Jetson TX2

Hi everybody,

I would like to optimize my retrained network mask RCNN with tensorRT to put it on a Jetson.
I know there are unsupported operations but can i optimize only a part of my network (the part corresponding to mobilenet for example) ?

Or there is another way to use mask-RCNN in real time on Jetson?

Thanks,

Hi,

I’m not aware of any official Mask-RCNN examples with TensorRT, however there seems to be a few options for getting Mask-RCNN on a Jetson.

  1. Mask-RCNN with Mobilenet meant for Jetson - https://github.com/gustavz/Mobile_Mask_RCNN
  2. Some models on Jetson with tf_trt - https://github.com/NVIDIA-AI-IOT/tf_trt_models/blob/master/tf_trt_models/detection.py

Perhaps the first link may be enough for your purposes without TensorRT, but you may be able to use these to get a TensorRT Engine for additional performance.

A general TensorRT workflow is TensorFlow -> UFF -> TensorRT+Plugin, and you can find some samples for each step in /usr/src/tensorrt.

Thanks,
NVIDIA Enterprise Support

Thank you for your answer.

  1. I have already see this repo. With this kind of solution (with tensorRT Engine), i obtain only 2fps…
  2. I’ll look into it. Is it possible to trained the optimized models with my own dataset? mask-rcnn-resnet for example ?

The use of UFF is only possible if all operations are supported, right?

Hi,

No, you can customize a layer that is not support by the parser.
See https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#extending

Thanks,
NVIDIA Enterprise Support