SLOW Semantic Segmentation on NVIDIA DRIVE Open Script

Hello,

I'm using a NVIDIA Jetson AGX Xavier

I'm trying the Semantic Segmentation on NVIDIA DRIVE Open Script: https://it.mathworks.com/help/supportpkg/nvidia/ug/semantic-segmentation-on-nvidia-drive.html

I've only chaned the

opencv_link_flags = '`pkg-config --cflags --libs opencv`';

to

opencv_link_flags = '`pkg-config --cflags --libs opencv4`';

because without it it doesn't compile.

There is another problem: Why the FPS are so slow? Mine goes at 0.39 FPS (as shown in the screenshot below)

I've checked the screenshot in the example and it goes at 8.73 FPS.

Thank you for your help.

Paolo R

Answer 1

[Hariprasad Ravishankar]

Hi Paolo,

Can you try setting the deep learning target library to TensorRT?

cfg = coder.gpuConfig('exe');

cfg.DeepLearningConfig = coder.DeepLearningConfig(TargetLibrary = 'tensorrt');

Hari

[Hariprasad Ravishankar]

In addition to this, you can also try the following to get a little bit more peformance.

1.TensorRT FP16 mode. Note that FP16 computation can result in lower accuracy from baseline FP32 computation.

cfg = coder.gpuConfig('exe');
dlcfg = coder.DeepLearningConfig(TargetLibrary = 'tensorrt');
dlcfg.DataType = 'FP16';
cfg.DeepLearningConfig = dlcfg;

2.You can use the nvpmodel tool to change the clock mode

For example:

sudo nvpmodel -m 0

See details here.