I'd be curious to know what kind of image recognition software the author used to detect relevant images and if it came with a significant performance hit.
I think you could do it very fast. It's just checking for the presence of static elements in the image, and you don't even need to check for every pixel. You would only need to decode the jpg and then check that maybe 30 pixels are the correct color (with a small tolerance). Also you know that every image is coming from an iOS device, so you can throw out any images that don't match a specific resolution, and you don't need to do any resizing or anything. I'm surprised that he didn't talk about resizing images on the device before uploading.
Pokemon go has many static elements in the UI (a pokeball front and centre being the most notable) -- I'd guess matching against a static template would be pretty fast. (Not my field, though.)