In the context of ongoing urbanization and warfare, accurate building damage detection is a crucial challenge. While most segmentation approaches rely on overhead satellite imagery, the task from ground-level perspectives remains underexplored. To address this gap, we introduce a dataset of 290 side-view images of Ukrainian buildings, manually annotated with six classes. We propose a deep learning-based segmentation system that divides images into fixed-size patches and augments them with global ConvNeXt-Large image embeddings and positional embeddings for encoding spatial location. The backbone follows a modified U-Net design, where the encoder is replaced by ResNet-50, SwinV2-Large, ConvNeXt-Large, Yolo11-seg, or DINOv2. We also evaluate a modified SegFormer-b5 for comparison.. We also study the effect of Felzenszwalb superpixel post-processing. Results show that U-Net with DINOv2 encoder and embeddings achieves the best performance on all six classes (IoU = 0.4711; F1 = 0.7462), while U-Net with ResNet-50 encoder and embeddings performs best on the three-class task (IoU = 0.7644; F1 = 0.8876). Embeddings strongly contributed to these gains: ResNet improved by +5.28 pp F1 and +7.81 pp IoU in the 3-class task, and DINOv2 by +4.71 pp F1 and +4.65 pp IoU in the 6-class task.
Lysanets et al. (Thu,) studied this question.