Sparse cloud good, Dense cloud = huge gap?

[khryst]

Hello everyone! First post here.

I've had decent successes with a novel setup . . . but I don't think the setup's the problem here. In one recent scan the cameras aligned successfully, and the sparse cloud looks like a great estimate of the object's shape. But a dense cloud is missing half the object, despite cameras aligning on that side and giving what should be adequate coverage . . . it's a failure I've not seen before . . . anyone seen similar and have any insight?

[Alexey Pasumansky]

Hello khryst,

Can you please post some screenshots and specify reconstruction settings used?

[khryst]

Good point, that would certainly help explain the situation 

Alas, I rage quit after fussing with it last night, turns out I had too many cameras in the mix. By removing half of them I got a much cleaner end result.

[khryst]

Okay, i re-attempted the build today. attached are the results

depth filtering mild, quality high

the sparse cloud has coverge on the entire outsole of the heel, and the top cameras should give it enough data to go by . . . but the dense cloud has that giant gap.

any insight? Thanks!

[James]

Nice shoe!

If I was trying to troubleshoot this, i would probably start by reducing the number of photos to ~50, and trying the dense cloud stage on low or lowest.

Align photos stage, which leads to the sparse cloud, works on an 'image feature' level, and the shoes have some lovely features, whereas dense cloud works at pixel level, so depending on texture scale of the object above a certain level of 'quality' (image downsampling) that pixel matching approach may fail.

[khryst]

so if you over-sample an object by using too high a Dense Cloud setting, it can introduce errors as it looks for details at a finer resolution than it should? . . . does that even make sense? 

[James]

So I take it you didn't try at a low setting, or you did and it made no difference?

If you think of it in terms of a laser scanner then it doesn't make sense, because the more densely you sample an object with a laser scanner then the better the representation will be.

With multiview reconstruction it makes sense, to me anyway, because densely sampling an image isn't enough on its own - you have to make correspondence with each sampled point in multiple images.

It's a bit like holding two 24x36" prints 1cm from your face - hard to tell if you're looking at exactly the same thing in each image, and depending on the scale of the texture you may be able to move the images around in front of your face without the view changing very much. If you hold them a bit further away it's like subsampling the image and the smallest things you can see now are more likely to be distinct from their surroundings.

If your images are perfectly sharp and the scale of texture is almost fractal like, i.e. as in aerial photogrammetry, then using ultra high quality and every last pixel makes sense.

If your images are not sharp, or the texture becomes less texturous at high magnification then there will come a point where higher 'quality' dense cloud building becomes pointless.

Imagine if you had a perfect 5 billion gigapixel camera and you could see individual atoms, the chances are they all look pretty homogenous and pixel matching wouldn't work. Similarly some kind of fake moulded plastic texture (not saying your shoes are fake or plastic!) might look interesting from a certain distance, but up close its just smooth plastic, which doesn't work well in photoscan.

I should point out I have no inside information whatsoever about the algorithms going on in photoscan, and most of this is based on what i've read on this forum and imagined when i should have probably been working.

[khryst]

James, that makes sense conceptually. Thanks for the follow up!

I only found success in this case (relative success anyway, I'm not truely happy with the result yet), by:

1. Reducing total cameras by 1/2

2. putting each ring of shots in it's own chunk

3. using Ultra High dense cloud setting

fewer cameras meant more accurrate alignment of the ones that were left, each ring of shots on it's own chunk gave more accurate dense clouds, and Ultra High dense cloud setting was the only setting that accurately reconstructed the "hollowness" of the shoe. The above advice to user lower setting usually results in a filled in shoe shape, without the gaps and voids inside.