NJIT
Toward learning-based reconstruction and part decomposition of man-made 3d geometry: neural implicit representations and scalable supervision
Abstract
dc:description.abstract<p>Digital three-dimensional (3D) models are central to engineering design, analysis, and manufacturing, but learning pipelines for man-made geometry often operate on sampled carriers that do not preserve all of the structure present in exact CAD representations. This dissertation studies learning-based reconstruction and part decomposition for structured man-made 3D geometry, from general object benchmarks to CAD-derived datasets, with a focus on neural implicit representations trained from signed-distance samples, point clouds, and tessellated meshes. The goal is to make these models more accurate, more part-aware, and more consistently supervised.</p><p>First, signed distance function (SDF) reconstruction with implicit neural representations is improved through periodic activation functions, positional encodings, and surface-normal guidance in an efficient auto-decoder pipeline. Evaluated on a widely used general 3D benchmark, these design choices improve the recovery of high-frequency detail and sharp features while reducing training cost relative to standard implicit baselines.</p><p>Second, a unified framework for CAD-derived geometry is introduced that couples curvature-regularized reconstruction with part-level segmentation in a single continuous field. By attaching a compact segmentation head to the reconstruction network, the framework jointly recovers geometry and coherent part boundaries from sampled carriers derived from CAD models.</p><p>Third, STEP-Parts is presented as a deterministic method for extracting stable part structure directly from B-Rep topology and analytic primitive information. Because the resulting partitions are invariant to tessellation choices, STEP-Parts provides reproducible geometric decompositions that support large-scale evaluation and reliable supervision for part-aware learning.</p><p>Together, these contributions establish a progression from general neural implicit reconstruction to CAD-native supervision, supporting more accurate and reproducible pipelines for reconstruction, segmentation, and downstream geometric learning on man-made 3D geometry.</p>
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy in Computing Sciences - (Ph.D.)
- Discipline thesis:degree_discipline
- Computer Science
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Fan, Shen
- 0009-0009-9364-5900
- Contributors dc:contributor
-
- Przemyslaw Musialski
- Jason T. L. Wang
- Ioannis Koutis
Subjects
dc:subject × 12Identifiers
dc:identifier.*- Repository record dc:identifier
- https://digitalcommons.njit.edu/dissertations/1879