The Doctoral School in Science and Engineering is happy to invite you to Kseniya CHERENKOVA’s defence entitled
AUTOMATED CAD (RE)DESIGN FROM 3D SCANS
Supervisor: Assist. Prof Djamila AOUADA
The automated conversion of 3D scans into semantically rich and editable CAD models, known as Scan2CAD, remains a critical challenge in reverse engineering. Solving Scan2CAD unlocks powerful capabilities for rapid prototyping of new designs, quality assurance of manufactured objects, legacy part replacement, and design analysis, ultimately bridging the gap between the physical and digital worlds with efficiency.
Traditional CAD modelling relies on Boundary Representation (B-Rep), a compact and structured geometry representation composed of parametric surface patches, edges, and vertices with defined topological connectivity. In contrast, 3D scanning typically yields point clouds or polygonal meshes, representations that lack semantic structure and are often plagued by inherent artifacts. Among those artifacts are the noise, incompleteness, smoothing of high-frequency details, that affect the geometric fidelity of a scanned object. This dissertation addresses critical limitations of existing Scan2CAD approaches, specifically their struggle with these scan artifacts and the complexities of real-world 3D scans, particularly when dealing with detail-rich geometries.
To overcome these limitations and facilitate data-driven solutions, CC3D (CAD Construction in 3D), a novel large-scale dataset of paired CAD models and realistic 3D scans, is introduced. Using CC3D, a series of specialized deep learning architectures are designed for realistic scan artifact reduction and accurate recovery of sharp geometric features. The developed approach enables topology-guided segmentation of 3D scans into B-Rep elements, yielding structured and continuous surface representations. The robust least squares fit algorithms for the primitives (line, circle, b-spline, plane, sphere, etc.) are introduced and analysed with respect to real-world B-Rep approximation. Notably, the approach achieves solid B-Rep structure reconstruction from million-facet 3D scans in under 3 seconds, demonstrating a substantial performance gain over state-of-the-art methods often evaluated on idealized synthetic data. Furthermore, this thesis introduces the SHARP benchmark suite, designed to foster standardized evaluation and accelerate progress in the Scan2CAD field. Consequently, this work underscores the necessity of moving beyond geometric fidelity to capture design intent, paving the way for future research focused on reconstructing not just shape but the design process itself.