Hi Ondrej,
I have uploaded a nice package with sample data and explanation 
EDIT :
I know the camera number and positions are not ideal, but im not sure that the low quality of the result is just that. as mentioned none of my tests didn’t make it notably better. so maybe there is a mistake somewhere else as well.
HI,
thank you for your data.
We’ve checked it with our developer and the biggest issue is, that the render is too perfect. Textures seems to be repetitive and there is no noise.
These are the images of different chairs, but they are considered as the same (the green lines should be for the same tie points).
Would it be possible to add some accidental noise to your render? Or use layers with noise images for geometry and texture images for colors?
Thanks for looking into it !
Hm, I can render with less quality, that introduces general noise, but not in the way that it would make a different place in the train recognizable with a different camera.
And yes, for finding the cameras it is very difficult as lots of images are looking similar or are partly or even totally identical. That is why the XMPs are so important, the camera specs, coordinates and orientation are known and super exact. To my understanding RC then should know exactly which images are overlapping and it shouldn’t be distracted by similar images in other positions in the train.
Generating special noise textures for different parts of the train to make them distinctable is a huge expense because it would prevent any instancing. That doesn’t seem to be realistic.
Hm and … I have made tests using only 20 cameras (the first icosaeder) … in this case there are no similar images and the output isn’t better as well.
I was thinking that something with the parameters of the XMP or in the XMLs generated by the script, or its commands are not ideal or wrong … that (at least a main) issue lies outside of the cams/images.
ps.
… wouldn’t it be a code issue that RC considers these images to be the same? their cameras are known and far from each other so it’s impossible that they show the same place in 3d … ?
if it is true what you’re saying, then wouldn’t that make using RC with rendered images almost impossible?
No, I’ve checked that with a tool developer. The features are important also with XMP’s workflow.
It is possible to use the rendered images, but they need to have “unique” texture. Or not repetitive or too perfect.
I suppose it was developed to scan mostly the real object, as the rendered one (but those are also used in some cases).
just to test that …
would it make a difference if there was an object in each seating compartment ?
like a red cube, green sphere, blue cylinder etc… just to know its not the same place.
or is it more small scale differences that matter ?
and ps …
it still sounds strange to me, if you remove all images+xmps from 0020+ and run the script on the folder, shouldnt the quality get better then? (it doesnt)
As you described the objects, they seem to be featureless to me (when they will have no pattern, only color). Not sure if this help.
No, as there is no good overlap between the images and still just few tie points are found.
hm but conclusion…
even if i would use a lot more cameras, maybe not from a central icosaeder but more a bunch at the windows pointing inwards, on the floor upwards, on the ceiling downwards.
that with perfect information in the XMPs.
then RC would still mix up images and wouldnt be able to make a better point cloud?
if that is the case then RC isn’t able to work with rendered projects at the moment.
(which would be surprising … i think there will be a high demand for this in connection with creating splats)
Hard to say without testing. I can imagine some parts reconstructed properly with good overlap and proper camera placing.
Also, when the render won’t be “perfect”, then it will be better for RealityCapture.
hm, ok…
i now had an idea that is worth testing …
if possible leave that box folder for a while …
maybe i can make an update in a week or two.
Sure, it is basically your folder.
ps. direct (email) contact with dev is not possible ?
I am sorry, but no.
ok easy,
i do my tests first and come back to you again w the results
thanks a lot to everyone involved
Hi, can I ask a follow up question on this matter ?
that must probably go to a dev .-)
the described workflow creates an XMP file
for each camera with the following rotations:
SplatCam_0000.jpg <xcr:Rotation>-0.999999506551129 0.000993427148363 0.000000000000000 0.000000368418214 0.000370855610632 -0.999999931232988 -0.000993427080048 -0.999999437784151 -0.000370855793631</xcr:Rotation>
SplatCam_0001.jpg <xcr:Rotation>-0.465313239383386 -0.885146083567306 -0.000000000000000 -0.244735505492486 0.128655227613827 -0.961016318674648 0.850639830719134 -0.447173616342797 -0.276491654921135</xcr:Rotation>
SplatCam_0002.jpg <xcr:Rotation>0.648965237845026 -0.760818059767741 0.000000000000000 -0.550608822698951 -0.469660230846094 -0.690108101624425 0.525046706907894 0.447856168309475 -0.723706299594201</xcr:Rotation>
SplatCam_0003.jpg <xcr:Rotation>-0.464098430497952 0.885783634308818 0.000000000000000 0.244696184492108 0.128206382205558 -0.961086312906789 -0.851314527131037 -0.446038649393104 -0.276248256362341</xcr:Rotation>
and then makes RealityScan export the camera information as CSV
using type Internal/External camera parameters :
#name,x,y,alt,heading,pitch,roll,f,px,py,k1,k2,k3,k4,t1,t2
SplatCam_0000.jpg,-8.794788820005311e-007,-15.50919752571237,2.650000182476623,-179.9430808077829,89.9787515277476,0,23.55132606606181,0,0,0,0,0,0,0,0
SplatCam_0001.jpg,7.825859921983813e-003,-14.15411175288523,2.647457104389309,117.7304962460894,73.94907320427725,0,23.55132606606181,0,0,0,0,0,0,0,0
SplatCam_0002.jpg,4.835945392382524e-003,-12.34588135372577,2.64334128245244,49.53636939952838,43.63866665685627,0,23.55132606606181,0,0,0,0,0,0,0,0
SplatCam_0003.jpg,-7.82600935515898e-003,-10.55411189894672,2.097457048062014,-117.6518896131663,73.96358409443978,0,23.55132606606181,0,0,0,0,0,0,0,0
I’m now trying for hours to replicate that by a script
to get the yaw, pitch, roll (CSV, type=Internal/External camera parameters) from the matrix (XMP)
is it maybe possible to get a tip ?
O:-)
( eg. left/right-handed change, rotation order, world-to-camera transform )
best, index
Hi @indexrefraction1
It is possible that there is not a connection between those values. It depends, which coordinate system you are using. As YPR can be in a local system and XMP is in XMP coordinate system.
I’ll send you one document about CSs in RealityScan.