Hi @jared, Ok, I wasn't interested in Clarifai that much until custom training became available, and now I'm fascinated. I have kind of a loaded/complex question I hope someone can answer. I hope I can be brief enough.
First of all, I've just gotten a chance to try this and I'm very excited about it! I work in a warehouse and we do a lot of very tedious inventory inspection that I feel could be automated. I'm just dreaming up ways of making use of this to actually make a dent in our inventory work.
Our inventory is mostly pails and boxes, in numbered racks. One of the limitations is there are no bounding boxes for what Clarifai finds, which is ok since I could probably just define a 'view' aesthetic concept where the racks are lined up properly in an image and then use software to cut it into squares, and inspect each rack that way.
Identification of what's in a rack so far is easy and works really well. Clarifai is easily trained to tell whether a rack is empty or not, whether it's pails or boxes, what brand/product line, etc. It works like magic and doesn't really take that many training examples! I love it.
What is daunting for me is when I try to go further than that, get into counting items, etc. I tried taking pictures of some 5 gallon pails in different quantities, and having a 'concept' for each quantity which is mutually exclusive to all the other quantities. Is this a valid approach? From what I understand, that is the only way to count with this kind of deep learning system?
I found it worked for awhile until I started getting more possible quantities. The error rate seemed to increase as I added more possibilities, to the point it was making ridiculous errors on quantities it once recognized ie mistaking 2 for 14. Do I just need a lot more training data because of all the mutually exclusive but similar looking tags I'm adding? The more I add, the more the error rate seems to jump on things it previously used to recognize correctly.
Is there any kind of rule of thumb for increased training sample data for an added number of 'concepts' that are similar but vary only slightly? Does it go up exponentially or something? Is the solution just to keep going and take lots and LOTS of pictures and tag all of them? Will it improve in this kind of situation eventually? I'm really interested to find out. Thanks!