neural_parts
neural_parts copied to clipboard
Comparison with primitive-based methods
From related papers like UCSGNet https://github.com/kacperkan/ucsgnet/ and DagAmendment https://github.com/eliemichel/DagAmendment, how do you relate the proposal to create procedural parameter constraints of a dag graph (union, subtraction, and intersection) to Neural Parts primitives.
The previous idea is to decompose 3d shapes into primitives so they can be adjusted by sliders.
Examples cases are like in a character creator to create new combination of shapes or via hyperparameter optimization as in DagAmendment to procedurally generate new models.

One idea you present is that everything is decomposed into a discs / spheres - genus 0. You morph a sphere into the part shape, disallow overlap and add a few other constraints.
References
https://en.wikipedia.org/wiki/Genus_(mathematics)