Breakthrough in 3D scanning means results are 4500% more accurate | Hacker News Reader

Breakthrough in 3D scanning means results are 4500% more accurate

lboro.ac.uk

61 pointsrbanffy30 comments

Here's the code: https://github.com/UoMResearchIT/Gryphon/blob/main/gryphon.p...

Based on a brief read, it looks like the "breakthrough" is discarding the two measurements in a point group that are the furthest outliers from the mean. That can't be it, can it? Because that's an afternoon's work for an intern, not a breakthrough.

Too late for you to hold up the train of progress, a patent was already filed. /s

You being the operative word, as in 21st century, patents are the definition of holding up progress.

That can be true, and so is the idea that if you could not reap the rewards of your inventions, you wouldn't invest the effort because it could only be a cost sink. It's more nuanced and contextual than that. That said, the patent / trademark (probably any similar system) has up and down sides

That's why I said "in the 21st century". As far as I studied this, patents are a really solid idea in theory, and perhaps they even worked as advertised at some point in the past. But the way they work today, they're thoroughly gamed - patent owners get the benefits, the society doesn't get its promised returns. But I agree, it's a pretty nuanced topic.

The idea of having time and protection to recoup your investment still applies today. Being "in the 21st century" doesn't remove this trade-off nor imply that progress is held up because of patents. It could be that progress suffers more without patents. Humanity and society works better with incentives and freedom of preferences

I'm not questioning the idea of temporary monopoly to allow recouping investment. The idea is prima facie good. But patents aren't just legally enforced dibs; the protection is granted in exchange for giving the invention to the public. This is where the patent system doesn't live up to its obligations in practice.

There's a bunch of problems that, together, make the whole greater (i.e. worse) than the sum of its parts:

- In many industries, but especially in everything involving computers, patent protection is too long - so long that the invention is likely to become useless before the patent expires. Technologies develop in lock step, riff of each other. If you lock down a crucial invention, you're also locking out a whole branch of R&D that may forever remain not explored[0].

- Patents are supposed to contain enough details in them that experts in a relevant field should be able to recreate the invention from the patent text. This is not what happens, we (the public) are being cheated here. The interesting patents are more like invention hashes - they contain just enough data to identify when someone is in the vicinity of invention, but not enough data to reproduce the invention itself.

- Patents are routinely awarded for things that do not have an inventive step (i.e. they aren't non-obvious). I'm willing to give some benefit of the doubt here, because ideas tend to look more obvious in hindsight, but I think we've all seen or heard examples of ridiculously obvious patents.

- Patents take too long to grant. The process between inventing something and getting it patented can take 2+ years - in that time, the patented invention may be well on its way to becoming obsolete (see point 1). This encourages shotgun patenting.

- Patents do not come with an obligation to actually productize an invention. Companies can, and do, patent inventions they have no interest in productizing, with the sole purpose of preventing competitors from exploring some technology or product space. This is as pure slowdown of progress as you can get.

I could go on. The point is, the ideas behind a patent system aren't bad - but the implementation decayed to the point of being just another rent seeking system parasitizing society.

[0] - It's something plainly visible in software, but perhaps software is a special case, due to existence of FLOSS. Maybe, if all software was proprietary, patents would only slow down progress. With FLOSS, nobody is waiting for patents to expire - eventually, someone will find a patent unencumbered alternative, release it for free, and the world ends up adopting that. And then ends up stuck with it, due to sheer momentum. So if that patent locked out the optimal solution to a problem, it's possible we'll never use it, because too much stuff uses the suboptimal, FLOSS version.

Yes, there are problems with the patent administration and enforcement, but in general they have a role to play in incentivizing. Saying it is only bad (rent seeking and parasitizing) is an oversimplification and exaggeration.

Could the case be made that certain FLOSS licenses (i.e GPL3) hold up progress by limiting how people can build on said software or deliver it to their customers / users? Or does the elasticsearch relicense in some way resemble a self-issued "patent" where they only enforce their rights against certain entities? (much like a patent holder has the task of enforcing their own IP and can pick and choose) And perhaps forcing you into choosing between buying their product or using a suboptimal free version?

Something can be remarkably simple and still a breakthrough at the same time. Handwashing, for example, is one of the greatest medical breakthroughs of all time despite its simplicity.

On the data side, a simple but effective algorithm is "lightweight coresets". A couple of line of NumPy, but strong guarantees and nice properties.

Lightweight corsets (shapewear using hose-like material) were also a simple, effective breakthrough.

It made Spanx into a ~$750M company.

Yes but telling people who already know how to wash their hands that they should wash them for “more than 1 second and less than 1 hour” isn’t novel, which is effectively what they’re trying to sell here.

There’s nothing in this “breakthrough” that isn’t already being done anyway, and the comparison stats are completely made up. 13.8cm resolution in current scanners? Have they even used a 3D scanner made after the millennium turned?

More accuracy may be better, but I’d like to see a quick method of extracting analytic geometry from a 3D scan.

We have laser scanners and Faro arms at work, and in the end it’s a person manually entering the geometry into Solidworks as the point cloud is useless.

If it was that stimple, then why it was not implemented and/or widely used before?

Almost every variation on strategies like that has been used with depth data. Median filters are more popular.

It was, but the authors have an incentive to pretend it's novel, the press has an incentive to pretend it's novel, and the patent office has an incentive to pretend it's novel.

The fact that the article mentioned 4500% three times, twice in the title alone, kind of warned me to take their news with a pinch of salt.

speaking as a humble statistician, If x is your sample, given x = {8, 3, 7, 1, 3, 9} sample mean = (8+3+7+1+3+9)/6 = 5.166, doctors mean = (1+9)/2 = 5, trimmed mean = (8+3+7+3)/4 = 5.25, median aka 50th percentile = (3+7)/2 = 5.

ofcourse programmers will have none of this. They'll code up a fancy function called computeMean(x) in java or python or what have you, with proper types & unit tests & suchlike, and declare "the mean is 5.166", and call it a day. No amount of persuasion will convince them that the mean is unknown, because the population's distribution is unknown, so population mean can at-best be hypothesized to some constant c, and then you can run a hypothesis test to see if you can disprove said hypothesis at some alpha level. ofcourse with a sample size of n=6 your power will take a beating and type-2 errors will spike, but who cares about these esoterica when our famed intern has coded up the mean function itself in one afternoon!

Snark aside, every summary statistic has widely varying properties. Sample means will asymptotically converge to the population ("true") mean as sample size grows. Sample means are an unbiased estimator of the population mean, so they have the mvue i.e. minimum variance, the variance decreases linearly with sample size. But that doesn't mean one should ditch biased estimators. Choosing a clearly biased estimator such as a trimmed mean instead of the sample mean, which is what they've done here, can definitely result in a huge breakthrough (at the expense of higher variance but that's ok).

Is this better than the commonly taught ransac (random sample consensus) for rejecting outliers?

https://en.m.wikipedia.org/wiki/Random_sample_consensus

> average margin of error for current 3D scanning machines is around 13.8cm

Sorry whaaat?? No 3D scanner I've ever used has a margin of error of 13cm!! Typically for person-sized objects margins of error can be expected to be 1mm typically.

A margin of error of 13.8cm in current 3D scans of human bodies? I think their baseline is bogusly large.

Yes that seems too big to be true

They're using a commercial scanner: https://www.sizestream.com/technology

The margin of error is less surprising if you consider that they're interested in the precision of measurements like hinged upper bust circumference, not the position of raw points in a point cloud. According to the paper "Erroneous measurements occur because of the 3 D Body Scanner’s incorrect placement of anthropometric landmarks or the physical positioning of body parts which increase measurement paths against the algorithms. The 3 D Body Scanner’s incorrect landmark and measurement placement, in this study, causes measurements with a range of up to 32.9 cm (M=14.6 cm; SD=12.67 cm)."

I.e. if the scanner confuses your arm for your leg, that's going to throw off the measurements a lot.

It seems like SizeStream could easily improve their product by applying some basic filtering to remove outliers.

Oh yeah. And a grainy surface would give a higher surface area which if you used that as a measurement for creating clothes would be bad, but less bad with improved surface measurements.

This is bad comedy. A 210-line-of-Python outlier remover is not a research advancement, it’s part of a grad student’s util file. Really reflects badly on everyone involved. It is partially explained that the work is published in Ergonomics, vs any computational conference - this reminds strongly of the “Tai’s Model” scandal.

https://academia.stackexchange.com/questions/9602/rediscover...

People have been doing submillimeter 3d scanning for a decade at least.

That looks like basic filtering by z score

Someone should also try median filtering, would that be considered another breakthrough?

Edit: just read the code, they also filter by the distance to the median.

4500%!!!!!!!!!!!!!!!!!!!!!!!!!!!11

That's 45,000,000 ppm!!!!!

In case it's missing, I think parent's tongue is firmly in their cheek - point is things like '4500%' are often manipulatively rendered that way to sound massive, vs. '45x' for example, which means the same, is just as massive, but perhaps doesn't sound as crazy?

Percentage are broken in many cases because they are ambiguous, and often difficult to compose. +50% then -50% is not identity but ratio 3/4.

A plain explicit ratio is sometimes better. Ratio in decibel, or "doubling number" can sometimes be valuable alternative, could be used more often.

+1 doubling -1 doubling is identity. Same for decibels.

Reply on news.ycombinator.com