This website requires JavaScript.

NAISR: A 3D Neural Additive Model for Interpretable Shape Representation

Yining JiaoCarlton ZdanskiJulia Kimbell ...+7 Marc Niethammer
Mar 2023
Deep implicit functions (DIFs) have emerged as a powerful paradigm for manycomputer vision tasks such as 3D shape reconstruction, generation,registration, completion, editing, and understanding. However, given a set of3D shapes with associated covariates there is at present no shaperepresentation method which allows to precisely represent the shapes whilecapturing the individual dependencies on each covariate. Such a method would beof high utility to researchers to discover knowledge hidden in a population ofshapes. We propose a 3D Neural Additive Model for Interpretable ShapeRepresentation (NAISR) which describes individual shapes by deforming a shapeatlas in accordance to the effect of disentangled covariates. Our approachcaptures shape population trends and allows for patient-specific predictionsthrough shape transfer. NAISR is the first approach to combine the benefits ofdeep implicit shape representations with an atlas deforming according tospecified covariates. Although our driving problem is the construction of anairway atlas, NAISR is a general approach for modeling, representing, andinvestigating shape populations. We evaluate NAISR with respect to shapereconstruction, shape disentanglement, shape evolution, and shape transfer forthe pediatric upper airway. Our experiments demonstrate that NAISR achievescompetitive shape reconstruction performance while retaining interpretability.