FloorLevel-Net

FloorLevel-Net recognizes and orders floor-level lines in street-view images to delineate adjacent building floors for spatial understanding in urban applications.

Key Features:

• Supervised Deep Learning Approach: Employs a supervised deep learning framework to locate floor-level lines in street-view images and address dataset limitations lacking geometric attributes or perspective priors.
• Dataset and Augmentation Scheme: Compiles a novel dataset and synthesizes training samples by integrating rich semantics from rectified facades with perspective priors derived from diverse street views.
• Multi-Task Learning Network: Utilizes a multi-task learning network that associates explicit facade features with implicit floor-level lines to enable precise identification and ordering.
• Height-Attention Mechanism: Implements a height-attention mechanism to enforce vertical ordering of floor-level lines and improve segmentation consistency.
• Geometry Post-Processing Stage: Applies a second-stage geometry post-processing step that leverages self-constrained geometric priors for plausible and consistent reconstruction of floor-level lines.

Scientific Applications:

• Urban Augmented Reality (AR): Provides accurate and ordered floor-level lines to improve context-aware image overlays in AR systems.
• Architectural Visualization: Delineates adjacent floors to support facade and building-level visualization tasks.
• Interactive City Navigation: Supplies spatial cues from ordered floor-level lines to enhance navigation and spatial understanding in urban scenes.
• Immersive Educational Experiences: Enhances contextual urban learning by supplying structured floor-level information for immersive visualizations.

Methodology:

Data compilation and augmentation by synthesizing training samples from rectified facades and perspective priors; network design implementing a multi-task learning framework with a height-attention mechanism to associate explicit and implicit facade features; second-stage geometry post-processing using self-constrained geometric priors to refine segmentation outputs and reconstruct floor-level lines.