in-depth comparison of the image preprocessing pipeline

[alexshepard]

this came up as followup todos during the bounding box experiments.

across the clades, evaluate the following different approaches:

• source photo squashed into 299x299
• source photo square center cropped to 299x299
• source photo center 85% center cropped, squashed into 299x299
• source photo center 85% cropped, then square cropped to 299x299

also, visualize the different approaches in a slide or presentation or notebook so that people know what we're talking about.

[alexshepard]

I made a comparison between three different preprocessing strategies:

1. no crop, just squeeze
2. central crop, then squeeze
3. resize with crop, no squeezing

[alexshepard]

here is how the "no crop, just squeeze" approach transforms an image

[alexshepard]

here is how the "central crop, then squeeze" approach transforms an image

note this is how we do things in production now

[alexshepard]

here is how the "resize with crop, no squeezing" approach transforms an image

[alexshepard]

here's a look at performance metrics across our normal categories for each approach.

based on this, I'm recommending we switch to using tf.image.resize_with_crop_or_pad to preprocess our images. I'll make a pull request for this in the inatVisionAPI repo.

[alexshepard]

I made a PR for this here inaturalist/inatVisionAPI#17 and requested that Patrick take a look