Floor Plan Image Segmentation Via Scribble-Based Semi-Weakly Supervised Learning: A Style and Category-Agnostic Approach

The field of architectural design is experiencing a transformative shift towards the integration of advanced computational methodologies, aiming to revolutionize traditional practices through automation. A pivotal aspect is the automation of floor plan recognition. This task faces challenges due to varied floor plan styles and the need for large-scale annotated datasets for learning-based methods, hindered by the lack of standardized visualization rules and specialized annotation knowledge. Our study introduces a novel scribble-based semi-weakly-supervised framework, merging weakly annotated and unlabeled images to boost model robustness and generalizability. This framework benefits from a simplified annotation process while retaining detailed information. Accordingly, we provide a new benchmark dataset for floor plan image parsing covering a wide range of architectural styles and categories. Experiments with our proposed framework demonstrate marked improvements in parsing accuracy and model adaptability, significantly surpassing current state-of-the-art solutions.