Reconstructing editable prismatic CAD from rounded voxel models

Reverse Engineering a CAD shape from other representations is an importantgeometric processing step for many downstream applications. In this work, weintroduce a novel neural network architecture to solve this challenging taskand approximate a smoothed signed distance function with an editable,constrained, prismatic CAD model. During training, our method reconstructs theinput geometry in the voxel space by decomposing the shape into a series of 2Dprofile images and 1D envelope functions. These can then be recombined in adifferentiable way allowing a geometric loss function to be defined. Duringinference, we obtain the CAD data by first searching a database of 2Dconstrained sketches to find curves which approximate the profile images, thenextrude them and use Boolean operations to build the final CAD model. Ourmethod approximates the target shape more closely than other methods andoutputs highly editable constrained parametric sketches which are compatiblewith existing CAD software.