Bachelor's thesis / 2024

3D Template Matching

A marker-less identification pipeline that turns CAD models into trained detectors for 3D-printed components.

Role
Research and implementation
With
THWS / Fraunhofer
Period
Aug-Dec 2024
  • Synthetic data
  • YOLOv11-OBB
  • Style transfer
Cover of the bachelor's thesis on marker-less identification of 3D-printed components

Overview

Industrial vision systems need substantial labelled data, but collecting it for custom 3D-printed parts is slow. This thesis built a modular pipeline that starts with the part's CAD geometry and produces the training data automatically.

The study also tested whether edge extraction and neural drawing styles could reduce the visual gap between synthetic renders and real camera images.

40
models trained
8
visual treatments
1,000
renders per object

Approach

  1. 01

    Generate

    Blender and BlenderProc place each CAD model in a controlled scene, vary pose, color, lighting, and camera position, then render 1,000 images per object.

  2. 02

    Prepare

    Known scene geometry supplies pixel-accurate masks and oriented boxes. The pipeline builds eight parallel datasets using color, edge detection, and drawing-style transformations.

  3. 03

    Train and compare

    Five YOLOv11-OBB model sizes were trained across the eight treatments, producing 40 models evaluated for accuracy, confidence, and inference time.

Object detections compared across color, Canny, HED, and anime-style transformations
Real-image detections across the tested visual treatments.

Outcome

Validation mAP50-90 remained around 0.97-0.99 across the study. On real images, anime, OpenSketch, and contour transformations produced the strongest weighted confidence, while simpler edge filters were faster but less robust.