Instance segmentation on Jetson Xavier NX using python

[akshay9]

I want to run mask rcnn on jetson Xavier NX. I was referred to the tensorrt repo to perform this. [https://github.com/NVIDIA/TensorRT/tree/master/samples/opensource/sampleUffMaskRCNN] However this only takes ppm images and also its in c++. I want to use mask rcnn segmentation on python on the jetso and also perform segemntation on frames obtained by a realsense connected to the jetson. Is there any other source for this? (https://github.com/NVIDIA/TensorRT/tree/master/samples/opensource/sampleUffMaskRCNN)

[AastaLLL]

Hi,

Below is an alternative:

1. Generate the TensorRT engine with the sampleUffMaskRCNN sample.
2. Inference the engine in python with OpenCV interface.
You can find an example below:

GitHub - AastaNV/TRT_object_detection: Python sample for referencing object...

Thanks.

[akshay9]

Hi,
Is this for object detection? As I am quite new at computer vision, can ou elaborate a little more as to what to do with th eengine adn how to go about it?

[AastaLLL]

Hi,

To inference a model with TensorRT, you will need to convert all the TensorFlow operations into TensorRT’s first.
Since the operations are not all one-to-one mapping, it needs some manually setting to wrap the operation.

As a result, please run the step.1 first.
It helps you to convert the operations into TensorRT API and serialize the output as a file.

After you got a TensorRT engine, you can deploy it with the sample shared in step.2.
It’s a python example to run a TensorRT engine with OpenCV image input.
And parse the output as the detection (bounding box+confidence).

Thanks.