How to train a new model for semantic segmentation

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Hi,

I am using Deepstream 4.0 on Jetson Xavier, there is a semantic segmentation model with 4 classes in “deepstream_sdk_v4.0_jetson/sources/apps/sample_apps/deepstream-segmentation-test/”.
My question is how do I train a new model with another dataset such as cityscapes?
Is there any tutorial or doc?

Thanks,
Jason

2

Hi,

The semantic segmentation model is TensorFlow-based UNet.
The .uff file of the converted model is located at /opt/nvidia/deepstream/deepstream-4.0/samples/models/Segmentation/semantic/

You can find some training / re-training / inferencing information on our NGC model page:
https://ngc.nvidia.com/catalog/models/nvidia:unetindustrial_for_tensorflow_16

Thanks.