I agree; looks like capturing the nose is very significant. Look forward to hearing about your progress.
I’m using 150-200 images per category (with 20% going into validation). I’ve been impressed with how well the CNN can resolve different perspectives of a given plane so far. I’m experimenting with gradually making the classifications harder. I would be really pleased/excited if I could eventually train a model to pick up some subtle visual differences between a C172 vs C182.
Hello, nice to know that you are working on this too.
The more subtle the difference are, the more images you need.
I’ve used 650 images of each classes for training and 300 for validation.
I think you need at least 100 of each classes to have a good validation set.
I think there is an option to not crop the images but resize it without preserving aspect ratio.
I have not rework on it since last week. Next week I will have more time to work on it.
For sure. I think it’ll be pretty cool to see it distinguish between the Cessnas - it’ll be kind of similar to identifying and distinguishing the Olsen twins from one another 
I’m gonna try and add more images to my training set, and hopefully things will get better.
Interesting - I’ll have to look at the resizing without cropping option. Might help here, although not sure by how much. Thank you 
In the previous library version it was croptype=CropType.NO in the tfms options.