search

feevos / resuneta

mxnet source code for the resuneta semantic segmentation models
Other
117 stars 30 forks source link

Training procedure. #10

Closed Akhilesh64 closed 3 years ago

Akhilesh64 commented 3 years ago

How do I replicate the training procedure you followed. I have the postdam dataset available with me. How do I preprocess the dataset to feed into the model.

feevos commented 3 years ago

Hi @Akhilesh64 , thank you for your interest in our work.

You have to: a) concatenate the optical with the DSM and create a 5 channel image R-G-B-NIR-DSM b) chop it up in training chips of size 5x256x256. In the file https://github.com/feevos/resuneta/blob/master/src/ISPRSNormal.py exist information on means/stds for normalization of the training chips. Chopping the full rasters into training chips can be achieved with code located in https://github.com/feevos/resuneta/blob/master/src/chopchop_run.py

I don't remember which DEM model we used. You may also find this tutorial a bit useful for understanding the data: https://www.linkedin.com/pulse/semantic-segmentation-tutorial-mxnetgluon-part-i-foivos-diakogiannis/

Hope this helps.

Akhilesh64 commented 3 years ago

Thanks this helps a lot.