Deep learning/computer vision project. Should I keep going?
Hi guys, I've been working on a deep learning algorithm to remove transparent elements on pictures. I'm starting to see some positive results: https://imgur.com/a/yEWkc
https://imgur.com/a/rXFnj
The results are far from perfect but if I compare to webinpaint .com, my method has some potential because it is using the background info instead of just removing it. It seems this type of editing is not too hard on Photoshop. I personally did not try. But looking at the results they get in tutorials, it looks like a relatively easy task.
I hesitate to continue working on this idea because my method is quite limited (it can only deal with transparent elements). Additionally the main use would probably be to strip watermarks. The legality of such tool could be problematic. Do you have any tips for me? Should I just open source the whole thing and move on?
Congratulations on creating an impressive application of deep learning! I think the educational value of what you are doing is in itself tremendous.
From a business standpoint...I see a lot of potential for this, but only you can decide what level of potential makes it worth jumping all in. Most of WebInPaint's traffic appears to come from watermark removal and they are #1 page 1 for Google on many key phrases for removing watermarks. This makes it easy for you to see a rough estimate of the traffic you could get if you entered the market and were able to dominate them.
If their level of traffic looks encouraging to you, then I would not worry that your method is too limited. In fact, I would be more concerned if you were trying to do too many things. Being the best watermark removal tool on the Internet would be better than being the 2nd best tool that also removes people from beaches. If you can be the best then you can charge a premium.
I would also not be concerned about the ability to remove watermarks using photoshop for 3 reasons. 1) Not everyone has the necessary photo editing software or skills. 2) Even if they do, your tool may be easier, or give them a headstart that saves time. 3) Your method "learns" a mask and reverses it, which is different than using a clone tool. Because of this, I would bet there will be cases where your approach is simply better than using Photoshop.
What's much more difficult to predict is where this could lead. Will your experience with automated image editing and access to customers lead to related tools that might be even more popular than this one? Will there be demand for an API? What other opportunities will appear that you don't yet know?
And, of course, there are negative projections that are difficult to predict as well. Will I be able to rank high in Google? Will I be able to sell a premium version or rely only on ad revenues?
Best of luck on making this tough decision!
Thanks for your insights. I think what will really make a difference is the quality of algorithm. Right now I don't have something that's good enough and getting it to work reliably could take a while. I don't know if it's a matter of working on it a few more weeks or spending the next 6 months. Maybe I will set a deadline and take a decision after.