Unable to get segmentation to work with Jetson TX2

Hi,

We have removed some non-essential layers which aren’t supported by TensorRT.
You can get more information here:
[url]GitHub - dusty-nv/jetson-inference: Hello AI World guide to deploying deep-learning inference networks and deep vision primitives with TensorRT and NVIDIA Jetson.

Thanks.

What’s a good tool to use for annotating images with shaded labels for training a custom image segmentation model?

Hi iandow,

Here are a few example open-source labeling tools for segmentation:

Depending on your dataset you may also be able to write a simple OpenCV tool to perform a rough labeling by color or texture, like is included with this drone dataset from the tutorial.

Hi there,
so now that the TensorRT support for deconv layers is out and decent I wanted to test it out with the segNet code.
So I used the original deploy.prototxt and changed the SEGNET_DEFAULT_OUTPUT to “upscore_21classes”. After the initial test I got something like this:

Which is ok I guess since I didn’t uncomment the bilinear interpolation I thought.

So I decided to check the classmap after the argmax in this part:

const int argmax = (c_max[0] == ignoreID) ? c_max[1] : c_max[0];

classMap[y * s_w + x] = argmax;

To see the output of classes, and probably this would resemble the output in Digits, but it didn’t, it was just the same pixelated output, you can see it clearer by checking these two sections:

That one is the output after the bilinear interpolation and this is the output after the deconvolution layer:

So my question is: what does Digits do differently? Since when doing the inference it gets something way smoother like this:

I guess it’s a small thing I’m missing out, so any help is appreciated, I’ll try to deploy my own network, probably it’s something regarding that, or could it be something in the argmax?

Edit, it worked with a different network, so probably that was it. I’m attaching a picture.

Hi, @JessicaHu. Could you tell me where is the example folder named FCN-Alexnet-Aerial-FPV-720p, please?