Lately I have been doing a lot of development work with turntables and what I have found is that RC is Terrible at calculating fixed camera turntable image sets.
Evert time that I get a failure its because RC generally resolves a perfect circle of cameras as an oval, or some images spiral off in strange ways.
I was wondering if we could get an option in the alignment step to allow for senarios where we know that every image is floating around a fixed point at stepped degree increments.
for context, my Images are shot in a black room with subject lit perfectly, 3 cameras vertucally stacked taking images at 10 degree increments, no blur or defocus.
The problem gets incredibly bad when scanning thin objects (Leaves) or smaller objects like pebbles.
Small objects, props, and nature elements.
The alignments generally fail on objects we have to either Flip (Rocks) or are thin, like Leaves.
Note that we have a black mask for objects we flip
that sounds strange. How much of your image does the object cover? It should be most of it, otherwise RC cannot calculate the distortion values since it needs all the geometric info it can get.
If you always use the same camera position and focal length and also are very careful about using the same angles (which you indicated), in theory it should be possible to set up a rig. I did not do that myself but I think it involves XMPs, where your camera positions are stored. There might be an option that you don’t use the camera positions as fixed but allow RC for small corrections.
When doing images of an object, flipping the object 180° and then doing more images you do not give RC enough information to make the connection between the image sets.
At the end, rotate the object again, but only 90° and do another round or two of photos.
That should give RC enough information to connect everything together.
Anywhere from 10% (a thin stick, small tall prop) to 60% (a small stone)
These objects are not big enough to fill frame without depth of field issues. and need to have 100% coverage, hense the flipping.
Typically shooting with a 100mm lens at F.16, the objects are not big enough to use anything wider.
We have never had any issues using Agisoft for this process, alignments come out perfect every time. however we want to use RC but RC just spews out unusable alignments half of the time.
'[XMP workflow]"
Sadly this cannot work as we have to adjust the setup for each object, our asset scale variance is all over the place. and one camera setup will not work for every prop type.
I have tried the draft XMPs in the past, with no success.
A interesting thing that happens quite often is that the DRAFT alignment works fine, a perfect circle of cameras is formed as you would expect. however doing a standard alignment will produce very strange camera positions
Note there is no background detail, the object is masked automatically.
@ShadowTail
This is how we do it for smaller objects,
Typically we have 3 cameras at varying heights, then doing a 120-180 flip. depending on the object. (for instance a stick will have a full 180 flip as you cannot shoot it lying flat)
One thing I have not tried yet is cropping the images so the object is filling 95% of the frame
hmm, I think somewhere between the relatively low coverage and the almost total lack of features lies the reason for your difficulties.
As to why the “other software” does a (seemingly?) much better job is up to speculation. I wished many times that some DEV would weigh in and shed some light on that so that people like us, who want to know WHY things happen (and not just how to work around problems) can understand RC better. My personal guess is that RC only aligns fewer images (in difficult sets, which I would say yours are) because it is aiming for a much higher accuracy which is neccessary for the higher LOD in the mesh. The “other software” is probably more forgiving at the cost of detail.
Something you might want to try is a textured background (distinctly different for each orientation of the object - no masks), then each should align well. Once you got the individual components, run an alignment with “align components only” set to true and hopefully the features on the object will be enough to pull everything together.
Cropping might be another way to go, although some other difficulties might arise from that. If, then crop all images in the same way.
Another thought: have you used lens grouping yet? That could just be the solution, since RC uses the info of all images to calculate one single distortion model. In theory, most parts of the frame should be covered somewhere and therefore giving you a better result. If you did not change any setting like aperture or focus in the whole session, then you are good to go. If you did change something, you might want to ungroup the images for a second alignment. Since RC uses the present components, it will very likely only adjust the camera orientation slightly.
And finally, did you try control points yet? Sometimes it can just make the difference. Start with one, somethimes that’s enough already. Be careful not to get trapped in the cycle of “just one more” though…
Oh, and I think what ShadowTail meant was to have circles in at least two, better three coordinate planes, not just in one.
E.g. if you scan a thin round pebble, dont just lay it on its flat side and then turn it upside down (180°), which would be most obvious and convenient. Also do one circle with it standing on the edge (you can use some modeling clay for that), ideally another one with it rotated 90° along the edge.
Forgive me if I am carrying owls to athens with that…
does a (seemingly?) much better job is up to speculation. I wished many times that some DEV would weigh in and shed some light on that so that people like us, who want to know WHY things happen (and not just how to work around problems) can understand RC better. My personal guess is that RC only aligns fewer images (in difficult sets, which I would say yours are) because it is aiming for a much higher accuracy which is neccessary for the higher LOD in the mesh. The “other software” is probably more forgiving at the cost of detail.