Doctoral Defence: Ahmet Serdar KARADENIZ

The Doctoral School in Science and Engineering is happy to invite you to Ahmet Serdar KARADENIZ’s defence entitled

Learning CAD Reconstruction from 3D Scans Aligned with Real-world Design Workflows

Supervisor: Assoc. Prof Djamila AOUADA

Modern engineering design and manufacturing rely on computer-aided design (CAD) systems to create and manipulate digital models of physical objects. CAD models represent not only the final geometry of an object but also the structured sequence of parametric features such as sketches, constraints, and modeling operations that define how the object is constructed and modified. While this representation enables flexible editing and reuse, such models are often unavailable for existing physical objects due to missing original designs or documentation.

3D scanning technologies provide a way to transform these objects to 3D point clouds or meshes. However, these representations lack parametric structure and are not directly editable within CAD software. Recovering feature-based CAD models from such scans (i.e., CAD reverse engineering) is fundamentally challenging, as the underlying construction process is not preserved in the final geometry and must be inferred from noisy scans. This requires reasoning over a large combinatorial space of possible sketch planes, parametric sketches, and modeling operations, all of which are interdependent.

Recent learning-based approaches have demonstrated the ability to infer CAD models directly from point clouds or meshes, enabling data-driven reconstruction. However, these methods primarily focus on recovering overall shape and often overlook the detailed structure of the underlying sketches that define parametric models. As a result, the reconstructed outputs are often misaligned with how CAD models are created and reverse-engineered in practice, limiting their editability and precision.

In contrast, industrial reverse engineering follows a structured and manual workflow. Starting from a 3D scan, designers analyze the geometry and extract cross-sections, which are planar slices that reveal 2D profiles. These profiles are reconstructed as parametric sketches and combined with modeling operations to rebuild the final CAD model. The gap between this workflow and existing automated approaches highlights a key limitation, namely that current methods do not explicitly model the intermediate representations that are central to CAD reconstruction.

This thesis formulates CAD reconstruction as the recovery of key intermediate representations underlying feature-based design. Specifically, it focuses on identifying cross-sectional planes, reconstructing parametric sketches with geometric primitives and constraints, and inferring the modeling operations that generate the final geometry. Since cross-sections reduce 3D geometry to 2D profiles, sketch parameterization becomes a fundamental step in the reconstruction process. In the first part of the thesis, this problem is addressed by formulating primitive parameterization as a set prediction task that infers unordered geometric entities and their parameters from raster images. This formulation is then extended to jointly infer geometric primitives and their constraint relationships within a unified architecture, enabling constrained CAD sketch parameterization. Inspired by industrial reverse engineering workflows, the next stage addresses 3D CAD reconstruction through a structured pipeline that detects sketch planes, recovers constrained sketches, and infers modeling operations to generate complete CAD models from 3D scans. Finally, real-world reconstruction is addressed by introducing a part-aware decomposition of multi-part objects and enhancing robustness to noise in scans. Overall, this enables the reconstruction of fine-grained, constrained, and editable models, advancing CAD reverse engineering toward more robust and automated real-world applications.

• Add to calendar