I'm mainly familiar with techniques for 2D image analysis and computer vision, e.g. CNNs. Are there specialized tools for 3D model analysis, or is it primarily a matter of machine learning as well?
keenan
@Kuragama Many techniques for 2D machine learning don't directly apply to 3D data, because the geometry is no longer on a regular grid; instead, different shapes might be encoded by different numbers of values, and each value might have a totally different meaning (e.g., a different coordinate of a different vertex of a different triangulation). You can use 3D grids instead of 2D images, but this is extremely wasteful in terms of memory/computation, and also is not coordinate frame invariant, e.g., two shapes could be related by a rotation and translation. (This can be dealt with by working with multiple rotated copies, say, but again this is quite wasteful in terms of memory.). These days people are starting to explore point clouds as a way to do learning on geometry, by always using the same number of points (say), but here you lose any benefit of adaptive sampling; you also may be forced to stick with operations that don't depend on a permutation applied to the points. And on and on. TLDR: machine learning on geometry is hard, but people are working on it!
I'm mainly familiar with techniques for 2D image analysis and computer vision, e.g. CNNs. Are there specialized tools for 3D model analysis, or is it primarily a matter of machine learning as well?
@Kuragama Many techniques for 2D machine learning don't directly apply to 3D data, because the geometry is no longer on a regular grid; instead, different shapes might be encoded by different numbers of values, and each value might have a totally different meaning (e.g., a different coordinate of a different vertex of a different triangulation). You can use 3D grids instead of 2D images, but this is extremely wasteful in terms of memory/computation, and also is not coordinate frame invariant, e.g., two shapes could be related by a rotation and translation. (This can be dealt with by working with multiple rotated copies, say, but again this is quite wasteful in terms of memory.). These days people are starting to explore point clouds as a way to do learning on geometry, by always using the same number of points (say), but here you lose any benefit of adaptive sampling; you also may be forced to stick with operations that don't depend on a permutation applied to the points. And on and on. TLDR: machine learning on geometry is hard, but people are working on it!