Deep Waltz: A Hybrid Neuro-Symbolic Vision Framework Using Waltz Labeling and Deep Neural Networks for Structural Scene Understanding

This work presents Deep Waltz, a hybrid neuro-symbolic vision framework designed to interpret 3D structural information from 2D line drawings. The approach combines a deep neural network–based junction classification module with a symbolic Waltz labeling constraint solver, enabling both perceptual robustness and globally consistent scene interpretation. The pipeline integrates image preprocessing, edge and line-segment detection, junction extraction, CNN-based refinement, and constraint satisfaction reasoning using legal junction label tables. Furthermore, an iterative EM-like feedback loop allows symbolic constraints to improve neural predictions over time. Experiments on synthetic polyhedral data, hand-drawn sketches, and real-image edge maps demonstrate improved accuracy, label consistency, and structural reconstruction quality compared to purely symbolic or purely neural methods. This work contributes a reproducible and interpretable neuro-symbolic framework for structural scene understanding and line-drawing interpretation.