CWC pre-trained model return wrong prediction on dataset image

[Guillaumebeuzeboc]

Hello,
First, thank you for the awesome work!
I am trying to run the pre-trained model of the CWC in the docker.
I downloaded the pre-trained model and froze it.
So I tried the inference on an image from the dataset (I tried different images).
And the result is not the one expected:

It seems that most of the background is detected as weed and that no crops are detected. I was expecting something close to the mask from the dataset.
I tried with tf and trt as backend and both return the same output.
From the readme we can see that the output should be way better. Any idea why I am getting this output?
Thanks,

Guillaume