Can old dogs learn new tricks?
We've been running clippingmagic.com since we launched it here on hn in May 2013 [0]. It was originally based on a purely "classical" algorithm, but since then Deep Learning has taken over the field, leaving us with the simple choice to adapt or die.
Most of our old competitors shut down. The new competition is "a bunch of kids". We're two 40+ year olds with young kids, mortgages, and a 15 year investment in classical image processing algorithms rendered largely obsolete by Deep Learning. Fun times.
So it's been challenging, interesting, humbling & frustrating to switch out our entire toolbox and go from zero to our first full end-to-end DL solution for background removal over these last few years. It's been a journey with fits and starts. Classical algorithms are slow-and-steady work, craftsmanship if you will. DL is much more "doesn't work, doesn't work, it's a miracle, no it's crap, fixed it with better data, now don't touch the golden recipe!"
(Along the way CM got a little DL boost, but it's a DL + classical algorithms hybrid)
The big takeaway for us has been that DL is all about the data. The networks get all the glory, but it's the data that makes the bulk of the difference.
While we have a host of ideas for changing the business model for background removal to enable a wider range of use-cases, right now we're primarily looking for feedback on this very early MVP.
Thanks!