Segmentation of Floorplans and Heritage Sites: An Approach to Unbalanced Dataset

Abstract

To identify structural deficiencies, floorplan digitization is essential for legacy buildings, and historically significant and culturally rich sites. However, these technologies heavily depend on vector graphics. For an efficient and economical solution, we present an approach with multi-loss functions to handle unbalanced classes in floorplans using U-Net architecture.

Keywords

• Deep Learning
• Unbalanced Dataset
• Semantic Segmentation
• Focal Loss
• Jaccard Loss