FloorLevel-Net: Recognizing Floor-Level Lines With Height-Attention-Guided Multi-Task Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Urban Computing

Background:

• Accurate identification of building floor lines is crucial for applications like urban augmented reality (AR).
• Existing street-view datasets lack sufficient geometric detail or perspective information for training floor-level line detection models.
• A significant data gap exists for supervised deep learning approaches in this domain.

Purpose of the Study:

• To develop a robust method for detecting and reconstructing floor-level lines in street-view images.
• To overcome the limitations of existing datasets by creating a new, augmented dataset.
• To enhance the capabilities of urban AR applications through precise building facade understanding.

Main Methods:

• A novel dataset was compiled and a data augmentation scheme was developed using semantic facade information and perspective priors.
• FloorLevel-Net, a multi-task deep learning network with a height-attention mechanism, was designed to associate facade features with floor lines.
• A two-stage process involving deep learning segmentation and geometry post-processing was employed for reconstruction.

Main Results:

• The proposed method demonstrated effectiveness in quantitative and qualitative evaluations on diverse datasets, including Google street views.
• The system successfully reconstructs plausible and geometrically consistent floor-level lines.
• Context-aware image overlay results showcase the practical utility and potential in AR.

Conclusions:

• The developed approach effectively addresses the challenge of floor-level line detection in street-view imagery, despite data limitations.
• The FloorLevel-Net architecture and augmentation strategy provide a strong foundation for future research in facade analysis.
• This work significantly contributes to advancing urban AR applications by enabling more accurate and context-aware environmental understanding.