Running Mask RCNN Inference

Hello there!

I’ve finished training a Mask RCNN model from TAO Toolkit Getting Started | NVIDIA NGC (skipped steps 6 and 7 of pruning and retraining) using a custom dataset on AWS EC2 and was then trying to run inference on my own Jetson Orin AGX.

I’m currently with a .tlt model that, running the !tao maskrcnn export function, exports it to a .uff file (despite the notebook saying it would generate a .onnx file… anyways)

The notebook then generates a TensorRT engine through the !tao deploy mask_rcnn gen_trt_engine function, that I’m not able to run on the Jetson since I guess it’s created based on the GPU architecture for the AWS machine that’s not the one on the Orin. So I also tried to run the !tao deploy mask_rcnn gen_trt_engine function straight on the Orin without any success.

Could you please guide me through these final steps… what should I do and then finally how to implement (using deepstream maybe?)

Thank you very much for any kind of help,
Best regards


Do you use PeopleSemSegNet?
If yes, please check below for a sample:

To create the TensorRT engine, please check the below page for the tao-converter:


Hi @AastaLLL,

I’m running the notebook from the link above, version 5.0.0 in notebooks/tao_launcher_starter_kit/mask_rcnn, the model is downloaded from !ngc registry model download-version nvidia/tao/pretrained_instance_segmentation:resnet50 (for what I saw in one of the links that’s also the architecture used on PeopleSegNet, right?)

I tried using tao-converter, but its documentation says I need an .etlt file - when running !tao maskrcnn export I get only a .uff as mentioned above. Any comments?

Thanks for the help so far!

You can copy .uff model into Orin and then follow TRTEXEC with Mask RCNN - NVIDIA Docs to generate tensorrt engine.

1 Like

Hi @Morganh, thanks for the reply!

I was able to export the model using:

and tried to run inference using without success… (ValueError: cannot reshape array of size 156800 into shape (1024,1024) - got this error when trying to run inference)

Any further ideas on how to make the engine run using python?

Thanks once again

You can refer to the tao deploy source code for Mask_rcnn.