Concerning Video About Pixel Scaling

[Agnaa]

This video isn't intended for a battleboarding context or audience, but I find some of the demonstrations in it troubling for this context. In short, people were using pixel scaling in high-quality real-world images to try to figure out the height/length of objects with known real-world lengths, and these estimations were off by 10-20%, even though the objects in question were only a few meters apart.

Until now, I assumed that pixel-scaling was to some level inaccurate, but that that error would only be on the order of 1-2%, unless the objects in question were noticeably closer/further away from the camera than each other. With this, it seems like the error is much larger and starts much earlier than I'd expected.

(As an aside, I think that angsizing already tries to account for this, so that shouldn't be relevant to this discussion)

Does this seem worrying to anyone else? Why/why not?

 

[Armorchompy]

no, not really

i mean all our calculations are approximations, nobody's claiming that our destruction values are anything but a huge simplification of reality, this isn't any different

also personal concerns probably aren't worth a calc group thread.

 

[Agnaa]

no, not really

i mean all our calculations are approximations, nobody's claiming that our destruction values are anything but a huge simplification of reality, this isn't any different

also personal concerns probably aren't worth a calc group thread.

Yeah they're approximations, but if we know that sizable issues can occur, and we can quantify them, shouldn't we work to rectify them?

Like, idk, penalizing any pixel-scaled values in works with 3-D visuals by 25%, getting stricter with rejecting pixel-scaling stacked over multiple images, or that measures objects significantly far away from each other? It'd be weird to say that, in fights, a character calculated to move at 300 m/s is faster than a character with a stated move-speed of 275 m/s, if we know that the 300 m/s value could be inflated, and in reality just be 250 m/s.

Where else would you want a potential issue with the site's calculation standards to be posted?

 

[KLOL506]

@DontTalkDT @Executor_N0 @Mr._Bambu @Spinosaurus75DinosaurFan @DMUA What do you guys think of this?

 

[Spinosaurus75DinosaurFan]

Personally I tend to try and deliberately low-ball in pixel scaling when there's uncertainty, for example scaling to an object closer to the screen.

 

[Armorchompy]

Like, idk, penalizing any pixel-scaled values in works with 3-D visuals by 25%

You're way oversimplifying things here, nevermind that it's a completely random number you made up. Remember that results can be decreased by perspective as much as they can be increased.

getting stricter with rejecting pixel-scaling stacked over multiple images, or that measures objects significantly far away from each other?

Yeah I guess we can just delete the wiki while we're at it no biggie

It'd be weird to say that, in fights, a character calculated to move at 300 m/s is faster than a character with a stated move-speed of 275 m/s, if we know that the 300 m/s value could be inflated, and in reality just be 250 m/s.

"It'd be weird to say a character we approximate to be faster than another character is faster than that character"

 

[Mr. Bambu]

Yeah I don't see an issue here.

Nobody's claiming our methods are perfect, nor is the video above proof that such shit is absolutely always a fail. Without watching it through, the cover says they calculated his height as 5'1 where his actual height is 5'6 iirc. That's five inches. Nothing to blow up established methods over, certainly.

 

[ElajRuengies]

This video isn't intended for a battleboarding context or audience, but I find some of the demonstrations in it troubling for this context. In short, people were using pixel scaling in high-quality real-world images to try to figure out the height/length of objects with known real-world lengths, and these estimations were off by 10-20%, even though the objects in question were only a few meters apart.

Until now, I assumed that pixel-scaling was to some level inaccurate, but that that error would only be on the order of 1-2%, unless the objects in question were noticeably closer/further away from the camera than each other. With this, it seems like the error is much larger and starts much earlier than I'd expected.

(As an aside, I think that angsizing already tries to account for this, so that shouldn't be relevant to this discussion)

Does this seem worrying to anyone else? Why/why not?

In the video Jesus Christ Cr1tikal specifically says

And they took this image here, and this is right after I had been hit with a fleshlight. So I'm slouched over, my knees are bent, and I'm leaning over the table, and then they started to make these calculations.

This is obviously where the discrepancy comes from. Competent pixel scaling takes posture and POV into account.

(Also, 5' 0.5" instead of 5' 6" is a 9.1% difference, not a 10% - 20% difference)

 

[Agnaa]

You're way oversimplifying things here, nevermind that it's a completely random number you made up. Remember that results can be decreased by perspective as much as they can be increased.

Yes, it's an initial suggestion.

I got the number by taking one slightly larger than the largest one demonstrated in the video.

I think results being increased by perspective is more of an issue than results being decreased by perspective, but if there is a way to only tackle inflated results I'd be happy to only apply it to those shots. I think those cases would be where the ruler is closer towards the camera and more in center frame than the thing being measured?

Yeah I guess we can just delete the wiki while we're at it no biggie

?????

"It'd be weird to say a character we approximate to be faster than another character is faster than that character"

When we know that one approximation has significant error bars that leave a real opportunity for it to be slower, and that one approximation doesn't, then yeah, it is weird.

Nobody's claiming our methods are perfect, nor is the video above proof that such shit is absolutely always a fail. Without watching it through, the cover says they calculated his height as 5'1 where his actual height is 5'6 iirc. That's five inches. Nothing to blow up established methods over, certainly.

The amount in absolute terms isn't very important, the amount in relative terms is. If there's a change in a linear measurement of 10-20% over a few meters, if that's stacked with any other pixel-scaling that compounds, and if it's used to get volume, that difference is cubed, giving a 33% to 72% difference.

I don't think making this approximation-method give slightly lower results to compensate is "blowing it up". It's just making the inaccurate approximation a little bit smaller to make it less likely to overstate things.

(Also, 5' 0.5" instead of 5' 6" is a 9.1% difference, not a 10% - 20% difference)

He also pixel-scaled a golf club, finding it to be 28 inches instead of 35. That's a 20% difference. Although I did make a miscalculation for that low-end, I inputted 5' 7" instead of 5' 6", getting me 9.83%, which I felt comfortable rounding.

In the video Jesus Christ Cr1tikal specifically says

And they took this image here, and this is right after I had been hit with a fleshlight. So I'm slouched over, my knees are bent, and I'm leaning over the table, and then they started to make these calculations.

This is obviously where the discrepancy comes from. Competent pixel scaling takes posture and POV into account.

There are three examples given in the video. Posture seems unlikely to apply to one of them (his knees are barely bent here, has a calc doing pixel scaling like that ever been rejected?), and physically cannot apply to another (It was done to a golf club). Although, to be fair, for that other shot, Charlie does say that the large shoes one of them is wearing could be contributing to it, but the person doing the pixel scaling took 20 pixels off their measured height to try to account for that (which visually looks like it does); seeming to show that even with the measures one may take to mitigate that sort of thing, it could still be off by 9-20%.

And hey, another thing that Charlie said was that PoV is nigh-impossible to take into account, since you'd need to know the camera's position and focal length. But if there is a good way to do that, I'd be all for it.

 

[Armorchompy]

I think results being increased by perspective is more of an issue than results being decreased by perspective

I don't, they're equal.

?????

You are suggesting huge ass sweeping changes and massive amounts of rules for a minor issue that happens in relatively few cases and barely affects results

When we know that one approximation has significant error bars that leave a real opportunity for it to be slower, and that one approximation doesn't, then yeah, it is weird.

Or faster. Don't forget about that. I don't think you understand how estimates are typically treated.

Anyway, literally no kind of math done in VSBW is exact, KE calcs often leave out elasticity or surface area that would affect even full impacts, speed calculations over a timeframe basically never account for possible acceleration or deceleration, lifting strength calculations don't account for the sort of grip and positioning of the one performing it which can make something like lifting a 50 kg weight a superhuman feat, our destruction values are a massively simplified approximation of how that stuff works in reality, the majority of speed calculations regarding non-video media require assumptions of timeframe, angular sizing is extremely fallible and can't account for some artists not having a great sense of perspective, basically every LS feat performed at super-speed would be massively increased by the tiny timeframe and so on. Pixel scaling is far from the most egregious case.

Bottom line is, we're not scientists and we don't do exact work. We're dudes on the internet who enjoy approximating how strong various feats are, and rate characters we like by those feats. Inaccuracies happen, we try to avoid them as much as possible but I guarantee or at least hope that not a single calc member here is under the impression that our results are objectively correct.

 

[Agnaa]

Maybe responding to this stuff will shed some light on my perspective.

KE calcs often leave out elasticity or surface area that would affect even full impacts

These both seem a function of our system being based off of energy rather than force, a switch which, if we had infinite resources, I'd like to be done, but it's way too big of a move for me to care about. Still, I hope that calculations along the lines of "Person falls out of a building and lands on a crash mat/truck full of pillows" are rejected, and I'd hope that KE calcs involving a character inside a vehicle only give the character durability from their own KE.

speed calculations over a timeframe basically never account for possible acceleration or deceleration

Last I was on the wiki, I heard we'd started taking that into account when evaluating flight speed feats. But in other situations, it seems nigh-impossible to determine and impossible to put reasonable limits on, and taking the average speed as their highest speed does work as a low-ball.

lifting strength calculations don't account for the sort of grip and positioning of the one performing it which can make something like lifting a 50 kg weight a superhuman feat

This seems nigh-impossible to quantify. But if it were semi-relibly quantified (say, by a table giving multipliers to the value based on which muscle group was used to lift it), I'd like it to be taken into account. For now, all that could be practically relevant is giving characters who lift sub-545kg amounts in ways that make that lift clearly superhuman, a superhuman rating. But that does seem like an extremely niche area. And really, not accounting for this just low-balls the characters.

our destruction values are a massively simplified approximation of how that stuff works in reality

Part of this is because of energy instead of force, and part of this is because more accurate methods aren't applicable to all the materials we'd like. I'd hope that if there were a more applicable method that it could be switched to, and I'd probably even be in favor of docking current methods if we know they're systemically unreliable by a certain degree.

the majority of speed calculations regarding non-video media require assumptions of timeframe

I'd hope that all of those timeframe assumptions are generously low, such that >98% of people looking at the calculation would think that the timeframe isn't too high.

angular sizing is extremely fallible and can't account for some artists not having a great sense of perspective

It seems nigh-impossible to determine whether an artist has a bad sense of perspective, or actually meant for something to end up a certain distance away. But if there are undeniable contradictions of relevance to an angsizing calc, I'd hope that it wouldn't be used.

basically every LS feat performed at super-speed would be massively increased by the tiny timeframe and so on

The same would apply to any AP feat, we don't let that happen since it inflates results to an unreasonable degree. But I would hope that LS feats performed at slow speeds would be penalized accordingly. Taking 10 seconds to stop a 5000 kg train moving at 10 m/s should not be taken as a Class 10 feat.

Pixel scaling is far from the most egregious case.

As I see it, the other cases are handled as best they can be. Some require far too much work, some are already low-balled, some require information that we can't properly gather from the text. But if you compare "being strict with pixel-scaling over many images, or over huge distances" to "destroying the wiki", that makes me worry that pixel scaling has actually gotten the least scrutiny

I guarantee or at least hope that not a single calc member here is under the impression that our results are objectively correct.

Some parts of our wiki's pages (the paragraphs below the destruction values table on the Calculations page) make me hopeful, while other parts (the introduction to the Calculation Guide page saying that the only part of calculations subject to debate are their assumptions, with the rest being objective inarguable facts) don't.

But if y'all still think that nothing needs to change for pixel scaling, then fair enough. I expect we've all pretty much run through our core arguments at this point.

 

[Armorchompy]

Some parts of our wiki's pages (the paragraphs below the destruction values table on the Calculations page) make me hopeful, while other parts (the introduction to the Calculation Guide page saying that the only part of calculations subject to debate are their assumptions, with the rest being objective inarguable facts) don't.

Here's the thing, past a certain point they are, if you pulverize a 1 x 1 x 1 meter cube of steel, under our system there is really only one possible result for that feat, and while that may not be fully accurate to reality, you can't really argue against it in our forums unless you have a better way somehow

But if y'all still think that nothing needs to change for pixel scaling, then fair enough. I expect we've all pretty much run through our core arguments at this point.

Yeah I don't think it should be, inaccuracy is inherent to basically all our math

 

[Antoniofer]

Although a range of error is espected in out calcs, we can do something to avoid increasing such calc. I myself have calculated have made several instances of multiple pixel scaling to obtain a result, but for some issues (in few cases, I made the calc but didn't uploaded the scaled picture) I was forced to repeat the calculation and get results with notable differences, not big differences, but in some instances led to tier jumps.

Taking into account several factors, such perspecive and drawing skill of the artist (if its a manga or comic), scaling more than one picture may lead to unaccurate results. Of course, I can't say that we ignore every calculation that uses more than one scaling picture, but if the final result is close to a tier limit, there's an elevated probability than the "true" result may actually fall in another tier, and I believe this is the reason why several popular verses get recalculations.

 

[Spinosaurus75DinosaurFan]

Well so the conclusion here seems to merely be more cautious in doing/evaluating calculations, I don't think any broad systemic changes are necessary.

 

[Hellbeast]

If anything wouldn't it be best to have evaluations be more strict?
Perhaps adding an extra evaluation to the process

 

[Armorchompy]

If anything wouldn't it be best to have evaluations be more strict?
Perhaps adding an extra evaluation to the process

Trust me, you don't want that. Think about how much it takes to get one (1) of us off our lazy asses and write "yeah sure" at the bottom of a blog, and consider how much longer it would take to pull that off twice.

 

[Hellbeast]

Eh fair point

 

[Agnaa]

Trust me, you don't want that. Think about how much it takes to get one (1) of us off our lazy asses and write "yeah sure" at the bottom of a blog, and consider how much longer it would take to pull that off twice.

just double the number of calc group members xd