Two thesis topics · Winter Semester · Application deadline: 10 Oct
M.Sc. Thesis (TUM × Bentley): Semantic Classification of Design Components in 2D Drawings & 3D Geometry (Un/Self-Supervised)
12.09., Studentische Hilfskräfte, Praktikantenstellen, Studienarbeiten
Two M.Sc. thesis topics in cooperation with Bentley starting Winter Semester. Focus: unsupervised/self-supervised methods for semantic classification of design components in 2D drawings and 3D geometry. Application deadline: October 10, .
Two Master Thesis Topics (TUM × Bentley Systems)
Start: Winter Semester
Application deadline: October 10,
Cooperation: Technical University of Munich (TUM) × Bentley
Advisors:
TUM: Panagiotis Petropoulakis —
Bentley: Georgios Pavlidis —
Possible examiners:
Prof. Dr.-Ing. habil. Alois Christian Knoll (Chair of Robotics, Artificial Intelligence and Real-Time Systems)
Prof. Dr.-Ing. André Borrmann (Computational Modeling and Simulation / Computing in Civil and Building Engineering)
Option A — 2D Drawings
Semantic Classification of Design Components in 2D Drawings Using Unsupervised Learning
Background. 2D CAD drawings and floorplans encode geometry and symbols (walls, doors, windows). Manual or rule-based parsing is brittle. Recent advances in self-supervised representation learning enable robust parsing without extensive labels.
Objectives.
1. Parse 2D CAD floorplans or raster drawings to extract candidate elements (lines, arcs, symbols).
2. Compute geometry- and context-aware embeddings (topology, adjacency, openings, annotations).
3. Use unsupervised/self-supervised methods (e.g., contrastive learning, clustering) to group elements into semantic classes (walls, doors, windows, columns).
4. Evaluate integration into design workflows and links to 3D BIM models for cross-modal consistency.
Expected outcomes. Prototype that identifies and labels basic components from drawings; analysis of methods; evaluation across datasets and design styles.
Option B — 3D Geometry
Semantic Classification of Design Components Using Unsupervised Learning on 3D Geometric Data
Background. BIM/CAD models store rich geometry, but semantics are often manually annotated or inferred via rigid rules. Geometric deep learning and computer vision can infer semantics directly from shape, topology, and context.
Objectives.
5. Analyze 3D geometry from BIM/CAD to identify doors, windows, walls, columns, etc.
6. Leverage unsupervised learning to discover patterns and groupings that infer semantic classes.
7. Exploit spatial/topological features (bounding boxes, adjacency, openings) to improve accuracy.
8. Assess integration into existing design workflows (e.g., component reuse, automated documentation).
Expected outcomes. Working prototype for geometry-based identification/labeling; comparative evaluation across datasets and contexts.
Downloads
9. Thesis PDF — Option A (2D Drawings)
10. Thesis PDF — Option B (3D Geometry)