Provide weights for segmentation network

[metzzo]

Hello!

Thank you for the interesting challenge!

I tried to reproduce your baseline method, but the weights are not publicly available. Training the network manually could be done, but complicates things for some people. For example, some do not have easy access to sufficient GPUs to train the model.

Would it be possible to share the weights of the semantic segmentation network with us?

Best regards,
Robert

[dzambrano]

hello! Thanks for the interest in this challenge!

While we're discussing about the possibility to have the pre-trained weights for the baseline, I would suggest to reduce the input resolution of the network. You can do that by reducing the MULTIPLICATIVE_FACTOR as:
INPUT.MULTIPLICATIVE_FACTOR = 1

You can check the config/defaults.py for details. This will reduce the input resolution but will keep the aspect ratio and generally the field of view.

Moreover, the baseline uses a pre-trained deeplabv3_resnet101; but you could consider to start from a shallower network.

[dzambrano]

A first baseline has been provided in the latest release v0.1.0