B.Sc. thesis · Graded 1.4 · 2025
Generating OWL knowledge graphs from object-detection output and evaluating how graph structure, format, and model size affect spatial question answering. Graded 1.4.
My thesis connects object detection with symbolic spatial reasoning. The generator turns bounding-box CSV data into OWL-compliant knowledge graphs, derives relations such as above, left of, and inside, and supports both single- and multi-camera inputs. I evaluated ten graph variants in OWL and plain triples across four language models.
Object detectors identify components and bounding boxes, but their output does not directly tell a maintenance assistant how those components relate to one another. Building that spatial knowledge layer by hand would be slow and tied to a specific machine.
Designed a deterministic pipeline that reads detected objects and pixel coordinates, derives spatial relations with geometric rules, and writes OWL ontologies with owlready2. For several camera views, it can either retain perspective-specific instances or summarize matching components into a smaller combined graph.
Generated and evaluated ten graph structures in two formats, producing 960 model answers across DeepSeek-R1, DeepSeek-V3, Llama 3.1 8B, and Qwen 2.5 3B. The results showed that graph complexity and serialization strongly affect smaller models; the thesis and defense received a grade of 1.4.
The Python pipeline reads object-detection CSV files, calculates bounding-box centers, and applies threshold-based rules for vertical, horizontal, and containment relations. owlready2 serializes the result as OWL, while a statement generator produces a compact triple-text alternative. Both formats were tested across single-view, multi-view, and summarized graph variants.
Algorithms that derive above, below, left-of, and inside-of relations from pixel coordinates.
Preserves separate camera perspectives or summarizes repeated components into a more compact graph.
Compares implicit and explicit relations, coordinates, single views, multiple views, and summarized views.
Reviews 960 answers for correctness and completeness across OWL files and plain triples.
Surveyed the state of the art in Semantic Web architectures and object detection.
Built the Python generator and the spatial relationship extraction algorithms.
Finished the quantitative LLM evaluation and submitted the thesis at TU Dresden.
Defended the thesis at TU Dresden; final grade 1.4.
Swipe or open to viewScroll, drag, or open to view
