Running code only with RGB images about kaggle_dstl_submission HOT 3 CLOSED

This is fine. There is no problems with running this solution only on RGB data. While IR channels are very valuable to differentiate greenery because of its special spectral characteristics, it is totally fine to use only high resolution RGB for any task. Basically, if human eye could see it, network will too.
However, different kind of images will require full retrain or at least fine tuning.

from kaggle_dstl_submission.

RGB in the Kaggle problem was in 11 bit (typically it is 8 bit), and it may happen that data that you are trying to run the inference at comes from a different distribution, and quality of the predictions will be low.

We were not allowed to use external data in that Kaggle problem, but if one does not have such a constraint, the easiest way to go would be to use external data like SpaceNet or extract images from OSM and train on it.

from kaggle_dstl_submission.

I was thinking about training the network with orthophotos, kind of higher quality images, but they are only RGB and as you mentioned they are 8bit with a resolution of 17500*12000. The positive think is that they are really geolocalized.
Is it enough to just change the num_channnel to make a prediction? I'm facing some issues.

from kaggle_dstl_submission.