This research paper introduces a U-Net-based network that uses spatio-temporal features with cross-attention transformers to extract information from historical maps. This method addresses the challenges of data-dependent uncertainty in the map processing methods, such as defects in the original map sheets and inadequate contexts when cropping maps into small tiles. The proposed model, U-SpaTem, performs better than other models that use either temporal or spatial contexts. The authors suggest that this method could also be applied to other fields with similar challenges, such as temporal sequences of satellite images.
Publication date: 20 Oct 2023
Project Page: https://github.com/chenyizi086/wu.2023.sigspatial.git
Paper: https://arxiv.org/pdf/2310.12616
