Is there a relatively easy way to get a colored voxel volume of an object in ShapeNetSem? I know there are objs, binvox files, textures, and the screenshots, but I haven't figured out a simple and effective way to convert such data to get w*h*l*3(for channels) matrices of the data.
The most effective way I could think of is voxel coloring using the turntable position screenshots and binvox files, but I am not so sure how far the camera is from the 3D object, at what angle it is at, and what sort of projection is done. And even I figured out how to get voxel coloring working, doing this on all 6k+ models would also likely be very inefficient.
I was also just trying to project the canonical positions (and ignore occlusions due to concave objects to simplify the problem a bit), but the size of the screenshots don't seem to be relatively similar. For example, for the front screenshot of model width id "2fa5e7c9e378b2589cd62032ff4f127f" is has a width of 15 pixels. From the top screenshot that same dimension has a size of close to 500 pixels.
I also asked this question related to this on Stack Overflow (because I was wondering if there was a way from the 3D geometry (obj + textures) instead of the binvox and screenshot files):
https://stackoverflow.com/questions/47386509/script-to-convert-3d-objects-that-are-textured-to-colored-voxels