As an example, for one branch, I took 49 photos total. 16 from an outer ring, 16 from a closer ring, and the rest were closeups (I tried to not move too much between shots). RC aligned 34 of them. Most of those that didn’t align in the main component were closeups. This makes some sense, as closeups of branches don’t really have many distinguishing features. I even had a hard time manually adding CPs (which I gave up on), as I couldn’t tell what was what.
As a result of the unaligned photos (I think), the resulting model has branch sections that are discontinuous (see below).
I’d like to improve my process. One critical thing seems to be to enable the closeups to align more often. As I’m only interested in the structure of the branch, I am able to add colored tags and otherwise to the branch sections in order to help RC align closeups.
So the question is: what would be the most effective “targets” I could add to the branch sections? Does RC use color in it’s matching algorithm? If so, then I could add colored tape to sections, or perhaps paint (or use a paint pen) to draw stripes around sections. This is what I am currently considering doing, but I’d love to hear your own ideas. Should I ensure that each photo has a minimum number of these targets in it’s view? If so, what might be a reasonable minimum?
Thanks in advance for your help!
Best,
Alex
Overview photo (there were about 16 of these):
Closeup:
Vertices of resulting model (high detail reconstruction):
What kind of resolution has your camera? Did you do a normal or high reconstruction?
The reason why the closeups won’t align is probably that you moved in too quickly. Even thoug the coloured stripes is a nice idea, it won’t do the trick in my opinion. Like with the circles, where you can’t change the angle too much at a time, you also need to “zoom in” slowly in small steps. You need to imagine all your cameras to form a web - only where it is connected will it align. If you do that, then RC will manage it even without fancy stripes!
Thanks Götz. I did attempt to take intermediate pictures as i moved closer and then around, but perhaps not enough.
So, overall, it sounds like what you’re suggesting is that in order to increase the quality of the model, I should focus on taking more pictures that form a tighter connection between successive shots, rather than using targets. That sounds like good advice… (thought I remain curious as to why you think the targets wouldn’t work… perhaps RC wouldn’t know apriori that there was just a single red tag, for example, and hence wouldn’t know for sure to match those two targets… that’d be a nice feature imho that could potentially reduce the need for extreme numbers of photos). The question then becomes, how to enable more photos without taking forever shooting each branch?
The challenge with this setup is that, when you rotate the branch on it’s rotating platform, you have to wait maybe 15 seconds or so for the branch sections to stop oscillating from the movement. One potential way of dealing with this would be to use a base equipped with a stepper motor to *slowly* rotate the branch, and stopping every so often (maybe 10 degrees?). That would keep the branch from oscillating, and would allow for more photos in a shorter period. (we don’t want to spend an hour taking photos - better to have each shoot done in 10-15 minutes if possible). Still not exactly sure how to move into the real closeups, as those seem to be what were failing… I can try just taking smaller steps I suppose…
Finally, to answer your specific questions, I was using a Nikon D7000 (4928 x 3264) in manual model, and using RAW output, and using high-detail reconstructions. I don’t think resolution is too much of an issue, as when you zoom in, the edges of the branch are spread across a number of pixels. So, if anything, I suspect that opening the lens up a bit (I was shooting at f22 to increase DOF) might help, as might a better lens (or perhaps even a full frame camera, but i’m not sure about that).
I think you’re on the spot with your analysis of the way RC works!
I fear that this is the price for the accuracy - other software is more forgiving but also less accurate.
To be honest with you, I don’t understand why you’re going through so much trouble with the turntable (which can also be square btw :-) - if you just set up the branch in one position and move around it (it will be quicker with some practice) then you can use the background to help with the alignment. And the wobbling time will fall away entirly along with the (almost 100%) certainty that they will stop at a tiny different spot, which will give RC unneccessary troubles.
I’m not a resolution fanatic but you can cetainly gain a lot by investing in that area. I used q 12 mpx LX100 for the last year. Now I got a X-T20 (24 mpx) with 10-24 zoom and the difference is mind-boggling. That’s probably also due to the much better sensor and the lens, but still.
I think f-22 was overdoing it a bit. What kind of sensor do you have? Because f-stops also are subject to adjusting them to sensor size like focal length. It depends on your lens, but most will be best at f-8 to -11 (below and beyond you get less sharpness).
I’m using the turntable because I really want to make sure I capture the fine details in the branch, and with a noisy background (this is all for testing methods that will be used in tropical forests and field stations), I’m concerned that the small branch ends will be lost. When they’re up against a high-contrast background, then it seems to me that there is a better chance they will be recognized as features. Having said that, perhaps it’s worth doing a trial of walking around vs rotating and seeing how it turns out.
Re: turntable shape, the top one can be square (that rotates with the object), but not the bottom one that sits on the floor and doesn’t move…
yes, definitely give it a try! I’ve been wondering how you would do that in the rainforest with the turntable and greenscreen - although there is a natural one there, hehe.
You might have a point with the contrast, but in my opinion RC excells in that over competitors and I am really not sure if it’s worth the extra effort. It will definitely work well if your coverage is good, that is if a branch has enough pixels to show a distinct texture. With photogrammetry it’s not about actual size at all, just about covering your target well.
Yes, I do use a tripod, but only inside. Outside, a monopod is entirely sufficient if it’s not close to dusk or dawn and in better light I even do it without any support. There is only the odd one that is not 100% crip, and that isn’t a problem at all if you have enough images close by. Although in the djungle it might be a bit too dingy.
Cool - I’ll give it a shot and let you know! btw, do you know if RC uses structure & color for its matching? i assume it does… just thinking of running a test with some color tags versus without…
I only know that it uses RGB info, I guess that counts as colour. But not in the way we use our eyes (or rather the brain) to tell if a picture shows the same. In our case, the difference in resolution can be enormous because we just fill in the gaps in our mind, But RC needs to identify certain features. I am pretty certain that those features are constitutet by a distinct distribution of colour changes in a certain area. This means that the same change needs to be identifiable in two different images, in our case with changing resolution. And then it makes sense that the steps shouldn’t be too big.
