Distinctive learning of latent space feature for occlusion-aware facade parsing

Significant strides in deep learning have propelled facade parsing in computer vision, a process that classifies architectural elements into semantic blocks. A key challenge is handling occlusions in facade images. Current methods struggle due to suboptimal use of multi-scale modules and insufficient differentiation of feature categories in latent space. Addressing this, we propose a novel multi-scale deep learning architecture, enhanced by a distinction loss function, to better capture multi-scale characteristics. This architecture includes a three-stream latent space feature enhancement structure: a main stream for primary processing and two auxiliary streams for feature refinement. Within the architecture, we utilize our designed dual-branch Context Aggregation Module to reconcile discrepancies between global and local features. We propose a distinction loss tailored to our network architecture, guiding the two auxiliary streams to concentrate on specific feature types, thereby enhancing discrimination and reducing confusion. On one hand, the proposed framework for distinct learning in the latent feature space introduces a novel learning paradigm for neural network training, where simple yet effective modifications to the loss function lead to performance optimization. On the other hand, the potential of our method to parse facades under occluded scenarios could be significantly impactful in engineering applications such as urban planning, architectural design, and autonomous driving. Our experiments on benchmark facade datasets demonstrate the superior performance of our approach in handling occlusions and effectively parsing facades, indicating the potential of our method to advance the application of facade parsing in increasingly complex scenarios.