Jason Antic

These are some really interesting test cases. Thanks for posting! Can you expand on the water/road confusion? Is there a particular picture you have in mind?

Roads and water look quite similar in the aerial with Westminster Palace (and in some places with the Crystal Palace surrounded by a large green area where the grass is colourized as light blue). I guess that's because they have similar shades of grey and texture details.

Ahh got it! So I'll just throw out some speculation here- As far as I can tell, the ImageNet dataset doesn't really have images shot at this perspective. I'm pretty sure part of the contextual information that goes into how the model interprets the scenes is relative (cont)

position and the general expectation that skies are up and water is down. I really think it boils down to that. So if you wanted it to perform well on images such as these, you'd want to supplement the dataset with aerial shots like this. I think that'd make a big difference.

Honestly it'd probably greatly improve the look of the buildings and other things as well.

That would be good. I will have a closer look on how to enrich the dataset. I'm also thinking that some data augmentation can be done by using different pairs of b&w and colour... cont

....e.g. Green/Red/Yellow Filter B&W etc as well as different vertical and horizontal distortions, vertical flipping transformations, multi cropping and then batch exported from Adobe Lightroom.

Awesome. What I'd actually suggest is any data augmentation where your altering images be done on the fly, as an additional transform in this list found in the ColorizeTraining.ipynb: extra_aug_tfms = [RandomLighting(0.1, 0.1, tfm_y=TfmType.PIXEL)]