Vectorizing historical maps with topological consistency: A hybrid approach using transformers and contour-based instance segmentation

Reducing the complexity of the workflow for historical map vectorization is essential to promote the widespread utilization of historical spatial data. Traditional pixel-wise segmentation followed by vectorization workflows suffer from tedious post-processing steps. To address this challenge, we introduce an innovative pure vector-based workflow. This workflow predicts object contours in vector format by assembling geometric primitives, such as line segments, in the correct order to form closed polygons. Consequently, the need for additional post-processing steps, typically associated with raster-to-vector data conversion, is eliminated. Furthermore, we have curated a publicly available historical map dataset called Sanborn-Vector, which holds significant potential for future research on vector-based historical map processing methods. To address the lack of suitable evaluation metrics for vector-based techniques, we have introduced a novel metric called structural panoptic quality (sPQ). This metric takes into account both the shape and positional accuracy of the vector output. Applying our proposed workflow to detect building instances from Sanborn maps has yielded simplified and intersection-free polygonal representations. We believe that our proposed workflow offers a fresh perspective on vectorizing historical maps, opening up new possibilities in this field.