vocksel
commented
3 years ago Coming back to this, the problems I notice most with shape detection is that:
- Circles get interpreted as squares a lot
- Squares/rectangles are too precise
- Chevron detection is too loose. You can essentially draw a half-circle and Geo thinks it is a chevron.
If we can tighten up these areas, shape detection should be a lot more reliable
Currently, shapes are not recognized as well as they should be, as pointed out by many uses on Twitter.
https://twitter.com/kevin1themorgan/status/1369436721968996364 https://twitter.com/ayameart/status/1369643717334859779 https://twitter.com/EugenicsOdyssey/status/1369595972037320706
I think this will mostly require tightening up the ratios we use, and limiting them to a specific range. If we are not within a range of certainty for a shape, it should be discarded completely. Worse case, we should return nil when we have no idea what shape the user drew.