seeme.ai - create your own image classifier without a single line of code.
A deep learning experiment
A few days ago, I started experimenting with automating computer vision with deep learning as much as I could.
I want people to be able to create their own image classifiers, but without:
- the need to find lots of images;
- knowing anything about deep learning or AI.
It is a long way from being generally usable, but there is no smoke and mirrors involved. It automatically collects the required data and serves up a pretty good deep learning vision model, no hands!
Right now, users can type the names of the things they want to classify and from then on Seeme takes over. It collects the data, trains and serves the model. Have a look for yourself:
Is it any good?
Whenever I train a new model, I automatically benchmark it against Custom Vision from Microsoft, and it seems to hold up pretty good. Their (awesome) v3 api is definitely stronger than the previous version.
That being said, there is no real tweaking going on at the Seeme end, as I have been focussing on getting it to work end-to-end. I hope to improve the training agent over the coming days…
Can I try it already?
I’m adding multi-user support at the moment, which is kind of a bare minimum to open it up…
Just about anything, I guess.
As I said, it’s a rudimentary end-to-end proof-of-concept that performes reasonably well.
Things I might add in the future:
- Manage the data;
- Share your model with others;
- Add actions on recognised objects;
- An app to snap and label pictures;
- More benchmarks;
A small thank you to:
the awesome people from fast.ai who taught me so much good things about deep learning, I look forward to every lesson…
If you made it this far, I would love to get your feedback: Questions, remarks, insults…