I'm curious about how much use these general-purpose computer vision APIs are actually getting. How many companies out there really want to sift through a lot of photos to find ones that contain "sailboat"? I'm inclined to think a lot more companies would want to find "one of these five different specific kinds of sailboats performing this action", which is definitely not among the tens of thousands of predefined labels that Google, and Amazon, offer with their general purpose models.
High-quality custom model training as a service seems much more compelling.
It's good for advertising. For example, Facebook analyzes all of your photos, and realizes there are a high number of images with sailboats.
That means you might have an interest in sailing, travel offers, or outdoor equipment, and Facebook can test that theory, and see if related advertisements have higher conversion rates.
It's also beneficial for user engagement, and by analyzing your photos, Facebook could recommend related groups in your area, or upcoming sailing events.
On the other end, imagine you make cosplay outfits for a living. You want to promote your business. Wouldn't it be efficient if you could only show your advertisement to Instagram or Snapchat users that post a high percentage of cosplay photos? That would result in a much higher click-through rate, and they could charge a higher premium for those targeted advertisements.
Or, what if you run a wallpaper site, where users upload wallpapers? You could automatically categorize and tag those images for users to search. Or, if someone is viewing a wallpaper of a sunset, you might want to show related sunset wallpapers that could interest them. That's pretty powerful. You could upload one million photos, and with a little work have them all nicely arranged in categories.
General purpose computer vision APIs are good if you're looking for breadth of concepts across many categories. For example, if you're Shutterstock and you're trying to make images searchable with very widely used, generally accepted concepts like "flower" and "car" then a general model would be good enough for you.
Custom computer vision models are good if you're looking for depth in certain categories. For example, if you're a gardening app and you want to take a pic of a flower and be able to recognize different species of flowers, then custom training is required.
There are some options with computer vision API companies where they will let you do custom model training. IBM will do custom training as a service for $$$$ but if you don't want to pay like crazy, Clarifai has a free (to a certain point) offering that lets you train a custom image recognition model on your own https://developer.clarifai.com/guide/train#train
It depends on the data set and the images in question.
Of course taking as an example for a sport’s clothing company where their digital assets are mostly related to their products a general purpose API might not get the subtle differences between two similar shoes, or clothing lines from different seasons. But it might be enough to help catalog that one image or video has a sports person in, and the other is a fashion shoot or product shot.
High-quality custom model training as a service seems much more compelling.