segmentation_demo doesn't work

[Benguigui__Michael]

Hello,

I am currently trying to perform cloud segmentation on 28x28 patches, by using "segmentation_demo.py" (open_model_zoo 2022.1.0).
Right before, I converted my ONNX model using "mo.py" (openvino 2022.1.0). I tested many paramaters at the conversion and inference level, but I can not get correct prediction masks as results (Outputs.zip attached). Any idea ?  I get similar results whatever the device (CPU, Myriad). I also attached:

- the ONNX model (model.LeNet_FCN.0.202205100731.28x28 ONNX.zip)

- the H5 model (model.LeNet_FCN.0.202205100731.28x28 H5.zip) converted to ONNX via "tf2onnx.convert"

- some input images (Inputs.zip) ...

-  ... with ground truth masks (GroundTruth.zip).

Thanks in Advance++

[Wan_Intel]

Hi Benguigui__Michael,

For your information, LeNet_FCN.0.202205100731.28x28 is not validated on Image Segmentation Python Demo.

The supported models for Image Segmentation Python Demo are as follows:

1. architecture_type = segmentation

    o  deeplabv3

    o  drn-d-38

    o  fastseg-large

    o  fastseg-small

    o  hrnet-v2-c1-segmentation

    o  icnet-camvid-ava-0001

    o  icnet-camvid-ava-sparse-30-0001

    o  icnet-camvid-ava-sparse-60-0001

    o  ocrnet-hrnet-w48-paddle

    o  pspnet-pytorch

    o  road-segmentation-adas-0001

    o  semantic-segmentation-adas-0001

    o  unet-camvid-onnx-0001

2. architecture_type = salient_object_detection

    o  f3net

Regards,

Wan

[Accelerator]

If I understand correctly, we can't use any custom architecture? Even 2 layers of convolution with an activation ? 

[Wan_Intel]

Hi Benguigui__Michael,

Thanks for your question.

This thread will no longer be monitored since we have provided suggestions. 

If you need any additional information from Intel, please submit a new question.

Best regards,

Wan