Convert to polygon inaccurate

Dear Team,

the “Convert to polygon” function does not work as accurate as I would need it to. When I get a proposal from the Instance segmentation assistant and would like to edit it I like to use “Convert to polygon” more than editing with the brush. But the conversion is not very fine grained and precision is lost. Using more points when converting would be great!

Thank you,

lucca

Hi cyFH3 and thanks for your feedback.

So if I understand this correctly, you want to control the fidelity of the conversion - and be able to set how “destructive” the conversion should be. Am I correct in that?

For me as a product person, the one thing I would love to get some input on is what metric we can use for this - should it be like “generate x polygons per 100px of the instance border” or something like that? It can be a tricky ropewalk between generating too few polygons (having some destruction of the annotation) and generating too many (making it impossible to edit).

Let me know what you think and I can sit down with our engineers and see what we can do!

Ps. You can also set the assistants to create polygons be default. If you go to advanced options (Project settings->advanced options) you can control switch between tools generating masks and polygons. Obviously, it doesn’t solve your current problem but might be nice in the future.

Alex Wennman
Cofounder

Hey Alex,

I am not sure if I want to control it. I just want the best possible outcome.

Concerning the metric I am not sure. I would think it would be good to use as many points as possible so that it is still editable (possibly only when you zoom in). But it is important how the points are set. Let’s say the area you want to cover is a square and you have four points to set then the result is quite different if you put the points on the corners of the square compared to if you put them in the middle of the connecting lines. I guess you would have to minimize the symmetric difference of the two areas. I don’t know how to minimize this efficiently (or if it’s possible).

When I set the assistants to create polygons or I set them to create masks and then convert the masks to polygons, will the result be the same?

Cheers,
Lucca

Hi again cyFH3.

Hmm - so I spent some time thinking about how we could automate this - the problem we will have is that how destructive we can be depends very much on the use-case. Sometimes, if you have simpler shapes, you can be really destructive and it will be fine. But if you are working with organic shapes (annotating a medical dataset, for example), any destructiveness at all can be a no-go. To that, would it be possible for you to send me one or two examples of what you are currently seeing when converting and what the ideal will be? (alex@hasty.ai)

Concerning the options question - yes, the result will be the same.

Kind regards,
Alex