Hacker News new | past | comments | ask | show | jobs | submit login

> Previous studies have found that dandelion seeds always have between 90 and 110 bristles, says Nakayama

Those poor research assistants. Imagine counting hundreds of dandelion bristles every day. Probably still not possible with AI/image recognition either.




That would 100% be possible with image recognition. Segment using UNet, then measure the remaining joined pixels.

(Well thanks for the downvotes. I do this as my job, so I guess I'm doing the impossible or something)


Perhaps people are wondering how your UNet is going to see the bristles on the other side of the dandelion seed.


Think you are misunderstanding what they are counting. The bristles of the seed sit on a stalk and are quite visible individually for the most part, because they are quite thin and stick out in different directions. If you were to place it so that the stalk points toward you, like in the Nature video, the bristles would stand out really clearly.

https://commons.wikimedia.org/wiki/File:Dandelion_seed_-_May...


Take photos from multiple angles?


And how do you deduplicate the ones that are visible on multiple angles?


Now you have a pretty ugly correspondence problem to solve.


I'm glad we have experts who can count to 100 instead. /s


Squash it flat.


I think the real question is, which is more cost/time efficient:

a) Hire someone or a company skilled enough to whip up some classifier that can count these from some accuracy, level, which entails paying a developer, and probably 5-10 graduate students to sit around and count dandelion bristles on a good number of dandelion seeds to create a training set.

or

b) Just have 5-10 graduate students sit around and count dandelion bristles on a good number of dandelion seeds

Given that a is a super set of b, and the extra portions of a are likely much more expensive than all of b, I think the answer is fairly clear...


Hm. Are NNs reliable enough to be sure you're getting a correct result?

I'd imagine this kind of task would lend itself to some old-school image recognition techniques - photograph against uniform background, threshold, mask the middle part and count contiguous regions.


ffs not everything needs ML. It's not like they need to scale to every individual dandelion in the known universe.


The bristles seem long enough. You could cut them, separate them, take a well contrasted photo, repeat. The image recognition task would be pretty easy at that point I think as you're just counting separated bristles.


That's kind of at the point where the solution takes so long to develop that you might as well count them manually.




Consider applying for YC's Spring batch! Applications are open till Feb 11.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: