CityGML Level of Detail 3 (LOD3) building model generation requires accurate façade representation. Existing workflows often rely on LiDAR, dense multi-view imagery, or pre-existing lower-LOD models, limiting their applicability in data-scarce environments. Additionally, most methods are developed on buildings with simple façade geometry, focusing mainly on detecting openings while keeping façade features planar. This study presents Omni2LOD3, a semi-automated workflow that extracts LOD3-ready geometric and semantic façade components from UAV photogrammetry and a limited number of omnidirectional images, challenging the assumption that façade-level reconstruction requires dense multi-view stereo or LiDAR acquisition. The study integrates existing methods and operationalizes them within a unified workflow for façade data extraction. The workflow is demonstrated on a building with complex overhangs, a curved façade, and partial occlusions. An LOD2 base model is derived from nadir drone imagery, while façade information is captured using omnidirectional images. Façade geometry is reconstructed through monocular depth estimation, and semantic features are extracted using color-based segmentation, establishing a lightweight baseline for semantic extraction from sparse image-derived point clouds. Evaluation against a terrestrial SLAM LiDAR benchmark indicates that the reconstructed façade geometry adequately captures LOD3-relevant structures, while HSV-based semantic classification results in 70% accuracy, reaching 81% after manual refinement. Overall, the results demonstrate the feasibility of recovering meaningful façade geometry and semantics from sparse omnidirectional imagery. These outputs constitute the LOD2 base model and semantically segmented façade point cloud that form the foundation for LOD3 generation, with full reconstruction currently under active development within the broader Omni2LOD3 pipeline.