Automated Detection and Encoding of Cartographic Signs in Historical Maps

Remi Petitpierre

cs.CVOct 9, 2026

Automated Detection and Encoding of Cartographic Signs in Historical Maps

Authors: Remi Petitpierre

Organizations: EPFL, Swiss Federal Institute of Technology in Lausanne. Digital Humanities Laboratory (DHLAB), School of Computer and Communication Sciences (IC). CM 1 468, Station 10, 1015 Lausanne, Switzerland. · EPFL. Laboratory of Urban Sociology (LASUR), School of Architecture, Civil and Environmental Engineering (ENAC).

Abstract

Symbols and icons in archival maps represent an expansive source of evidence for historical geography. In this paper, we present a two-stage framework for (1) detecting cartographic signs in historical maps and (2) generating interpretable sign embeddings that capture both semantics and stylistic variations. We also introduce CaSiDD, an open benchmark dataset spanning multiple sign categories and diverse cartographic styles, designed for generic sign detection. Using YOLOv10-X, we demonstrate the practicality of automated sign detection on historical maps. We then present strategies for creating contrasted, informative embeddings of cartographic signs. We approach embedding design systematically, maximizing expressiveness while suppressing digitization or background material confounders. Our final method, based on an adapted RADIOv4 backbone, effectively encodes both sign form and semantics, while producing interpretable representations. This work lays a foundation for scalable evidence extraction for the comparative study of map signs within cartographic culture and geohistorical research.

Figures & tables

Figure 1: Illustration of the challenges to the computational assumption of sign discreteness and the pictoriality of the representation. Signs may overlap (e.g. Fig. 1 c), exhibit ill-defined, ambiguous boundaries (e.g. Fig. 1 a, 1 d), or appear utterly pictorial (e.g. Fig. 1 b).

Figure 3: (left) Synthetic and pseudolabeled data used for training the foreground segmentation model. Synthetic data are produced by overlaying digital cartographic signs onto map fragments. Pseudolabels are created from real sign samples using traditional computer vision techniques. Synthetic data are associated with perfect masks, while pseudolabels help the model generalize to more realistic sign samples. (right) Inference results. A soft foreground mask is predicted by a neural network, and background content is removed based on the smooth mask. The foreground segmentation model erases undesirable background content quite effectively.

Figure 4: Two-dimensional, constrained t-SNE visualization of the sign embeddings, derived from an iconographic map of the Canton of Zurich, ca. 1770 Walser and Homann’s heirs (1770) . (a) In the base model using RADIOv4, spurious clusters appear (e.g. at the top), and background color heavily affects spatial structure. (b) In the adapted model, clustering is more shape-dependent and distinctive.

Model	Treatment	Precision [95% range]	Recall [95% range]
DINOv2	base	82.5% [81.0–83.9]	83.4% [82.1–84.6]
	grayscale + foreground seg.	77.3% [76.1–78.7]	78.4% [77.2–79.7]
	adapter	81.1% [80.0–82.2]	81.9% [80.9–83.1]
DINOv3	base	79.5% [78.3–80.8]	80.6% [79.5–81.7]
	grayscale + foreground seg.	74.4% [72.8–75.7]	76.3% [75.1–77.5]
	adapter	78.3% [76.7–79.7]	79.3% [77.5–80.6]

Table 1: Performance metrics for the classification of icons into 24 semantic classes. In each experiment, classification uses the MLP classifier. For both precision and recall, the reported ranges give the 2.5th–97.5th percentiles across 100 trials. The confusion matrix and the detailed per-class performance are provided in Figs. 5 and 6 .

Figure 6: Precision-Recall (PR) mean of the sign classification task for each semantic class, computed from the embeddings generated by the adapted RADIOv4, using the MLP classifier and 24 classes. The reported results aggregate 100 trials, each with

n_{val}=3,641

validation signs. The vertical bars denote the 2.5th–97.5th percentiles across 100 trials. Labels are ordered by their frequency in the training set (from

n_{1}=6,679

for trees to

n_{24}=4

for battlefield). The mean precision and recall values are reported in Table 1 . Plot (a) reports the original PR-Mean for each class, where PR-Mean = (Precision + Recall) / 2, whereas subfigure (b) presents the PR-Mean divided by the natural logarithm of the training set class counts. The horizontal bar indicates the mean of the tree class. For example, grass and vine signs are classified well, especially considering their frequency in the training data.

Treatment	Precision [95% range]	Recall [95% range]
color background (base)	38.8% [36.2–40.1]	51.6% [51.1–52.0]
tree class baseline	25.6% [24.9–26.2]	50.6% [50.6–50.6]

Table 2: Impact of weak features on sign classification. Classification uses logistic regression. Scores use fixed full-corpus weighting by class frequency. For both precision and recall, the ranges give the 2.5th–97.5th trial percentiles. The last row is the constant tree class baseline.

Model	Treatment	ROC-AUC [95% CI]	wCM@1 [95% CI]	wCM@3 [95% CI]
DINOv2	base	74.9% [73.3–76.4]	63.2% [60.3–66.2]	50.2% [47.3–53.1]
	grayscale + foreground seg.	73.6% [72.1–75.1]	59.1% [56.0–62.2]	45.9% [43.1–48.8]
	adapter	75.3% [73.7–76.8]	59.1% [56.1–62.1]	45.9% [43.1–48.7]
DINOv3	base	75.6% [74.0–77.3]	65.9% [63.0–68.7]	52.2% [49.1–55.2]
	grayscale + foreground seg.	74.1% [72.5–75.7]	60.6% [57.5–63.7]	48.0% [45.1–50.9]
	adapter	76.5% [74.9–78.2]	62.5% [59.5–65.5]	49.7% [46.7–52.7]

Table 3: Performance of the stylometric experiment in each setting. ROC-AUC measures the contrast between within- and between-community pairs. wCM@k indicates the within-community matching rate, i.e., the average proportion of nearest neighbors sharing the same community label for any item. CI denotes the Student-

t

95% confidence interval of the mean across 85 community batches.

Model	Treatment	Background color ↓	Foreground color ↑	Shape (HOG) ↑
DINOv2	base	15.9%	7.9%	23.5%
	grayscale + foreground seg.	8.7%	8.5%	32.3%
	adapter	5.2%	4.4%	30.9%
DINOv3	base	17.5%	15.5%	9.8%
	grayscale + foreground seg.	12.2%	13.7%	20.2%
	adapter	11.9%	17.8%	19.2%

Table 4: Ability of the embeddings to describe graphical differences in RGB color and shape. Each value indicates the linear association between distances in the vector spaces, measured on

8\cdot 10^{6}

sampled distance comparisons using Pearson’s coefficient. The best scores are highlighted in bold. For background color, lower values are preferable (↓).

Historical Document Processing

Historical document processing uses AI to transcribe, structure, restore, and analyze manuscripts, books, newspapers, and archival records.

See all 76 papers

Generalizable Multiscale Segmentation of Heterogeneous Map Collections
Mar 5, 2026Remi PetitpierreSemantic Segmentation Image Segmentation
EPFL, Swiss Federal Institute of Technology in Lausanne. Digital Humanities Laboratory (DHLAB), School of Computer and Communication Sciences (IC). CM 1 468, Station 10, 1015 Lausanne, Switzerland. · EPFL. Laboratory of Urban Sociology (LASUR), School of Architecture, Civil and Environmental Engineering (ENAC).
Deep learning enables urban change profiling through alignment of historical maps
Feb 2, 2026Sidi Wu, Yizi Chen, Maurizio Gribaudi +4
Institute of Cartography and Geoinformation, ETH Zurich, Switzerland · CRH, École des hautes études en sciences sociales (EHESS), France · Photogrammetry and Remote Sensing, ETH Zurich, Switzerland +1
SignSeek: Learning Transferable Representations for Sign Dictionary Retrieval
Sep 3, 2026Sobhan Asasi, Ozge Mercanoglu Sincan, Richard BowdenContrastive Learning Representation Learning
CVSSP, University of Surrey

Mar 5, 2026J/K move · Enter open · S save

Generalizable Multiscale Segmentation of Heterogeneous Map Collections

Remi Petitpierre

EPFL, Swiss Federal Institute of Technology in Lausanne. Digital Humanities Laboratory (DHLAB), School of Computer and Communication Sciences (IC). CM 1 468, Station 10, 1015 Lausanne, Switzerland. · EPFL. Laboratory of Urban Sociology (LASUR), School of Architecture, Civil and Environmental Engineering (ENAC).

Semantic Segmentation Image Segmentation Domain Generalization

Open paper AskPDF arXiv

Historical map collections are highly diverse in style, scale, and geographic focus, often consisting of many single-sheet documents. Yet most work in map recognition focuses on specialist models tailored to homogeneous map series. In contrast, this article aims to develop generalizable semantic segmentation models and ontology. First, we introduce Semap, a new open benchmark dataset comprising 1,439 manually annotated patches designed to reflect the variety of historical map documents. Second, we present a segmentation framework that combines procedural data synthesis with multiscale integration to improve robustness and transferability. This framework achieves state-of-the-art performance on both the HCMSSD and Semap datasets, showing that a diversity-driven approach to map recognition is not only viable but also beneficial. The results indicate that segmentation performance remains largely stable across map collections, scales, geographic regions, and publication contexts. By proposing benchmark datasets and methods for the generic segmentation of historical maps, this work opens the way to integrating the long tail of cartographic archives to historical geographic studies.

Abstract

Historical Document Processing

Explore similar work

Generalizable Multiscale Segmentation of Heterogeneous Map Collections

Deep learning enables urban change profiling through alignment of historical maps

SignSeek: Learning Transferable Representations for Sign Dictionary Retrieval