Hello! We are making food photogrammetry with RealityCapture and having troubles aligning images. We are using turntable Orangemonkie Foldio360 with a camera Canon EOS 5D MarkII and a light box. Our imageset consist of 3 sessions (top, middle, front). Each session has ~36 images (10 degrees angle shift). We have problems aligning cameras for some of the dishes. We tried playing with different aliment settings inside the program as well as adding anchor points to the images. In some cases points help to get good aliment but sometimes not. We don’t use masks to hide turntable elements but during RAW image processing we are cutting out all possible background elements that can have negative effect to the feature detector. I will attach thumbnail with some of the photos from the set that can’t be aligned good.
Maybe someone can give us advices that could fix this aliment issues. We read a lot of RealityCapture forums about aliment issues and tried a lot of different possible solutions but nothing of that didn’t help so far. Thanks in advance!
I had a problem once with a turntable setup. Mine was much less professional than yours Maybe my problem will help you with yours.
The problem that I encountered was that RC was finding more tie points in the background than on the object of interest. This fits in with ShadowTail’s commend about insufficient object coverage. Although I have had perfect alignment results before with much less than 80% coverage, more is definitely better.
I know you said that you’re removing background elements, but that may not be sufficient. Have you tried checking for match points across images that do align? (Alignment > Analyze Tab > Show Matches).
You can see if the match points show up mostly in the background or on the object. At least that’s what helped me figure out the problem.
Also, what exactly are you doing to your RAW images?
I want to add that the ceramic plate is a nightmare for photogrammtetry and I would bet that there are hardly any features at all. Somebody has made a turntable out of a plate for exactly that reason! Although that one was also opaque…
If you want to get anything on the plate you need to give RC a non-reflective pattern to latch on to. The best way to achieve that is to cover it in a thin, irregular coat of paint or similar. If you re-use the plate for several projects you could also model that and add it later to the content (food).
ShadowTail, The problem with 80% coverage is that we can’t move camera closer to the dish because it’s restricted by the tripod. Also we can’t zoom closer to the dish because I know that it’s restricted to change focal length when you do shots for photogrammetry.
Tim B, yes we tried checking Matches and images and the most number of points are on the food. I’ve attached screenshot of that. We convert (to JPG) and adjust RAW
images in Adobe Lightroom. There we adjust white balance, exposure, contrast, remove specular highlights and glossy effects as much as we could. Also in Lightroom we crop images to cut out all unnecessary parts like turntable elements, and other background stuff.
UPD. We created new RealityCapture project and tried to align the same imageset with default setting and we got 112/112 cameras alignment. The strange thing is that the settings on that previous project was the same (screenshot below).
it’s not really surprising since especially with difficult image sets RC displays quite some randmoness. I’m glad that you got a good alignment. Just make sure that there aren’t any “accidents” (as in grossly misaligned cameras) in the component, which is what sometimes happens in those cases.
Zooming is no problem at all if you keep it within reason and especially if you have a good quality lens. If you do all your shots at one zoom setting without changing anyhting, then it is as good as with any fixed lens. The sole difference is that the geometrical complexity is usually much lower with fixed lenses. If you have a zoom lens, there is no difference if you have it at one of the extremeties or somewhere in the middle. It’s just that the ends are the only reliably reproducable positions. You can even use different zoom settings, as long as you make sure that there are enought images with the same position. I would think that 3 is the absolute minimum but the more, the better. And this will also only take effect if you use exif grouping…
Tim B, Now we try to cut out (or paint out) the plate in Adobe Lightroom for the cases when it has to much glossiness. But afterwards we model it from the scratch and bake shadows and other effects in the Substance Painter. In most cases this solution helps to align images better but there are still some cases when alignment is impossible with this solution. I’ve attached few screenshots for such example.
Maybe my problem will help you with yours.