Segmentation of historical maps without annotated data - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Communication Dans Un Congrès Année : 2021

Segmentation of historical maps without annotated data

Jean Camillerapp
  • Fonction : Auteur
  • PersonId : 932492

Résumé

This paper presents the method which we submitted to the competition of Historical Map Segmentation, in ICDAR'21. The goal is to segment document images of Paris maps from the beginning of the 20th century: delineate the content of the map and locate the graticule line intersections, which give indications on the geographic coordinates of the map. Our contribution is entirely led by a rule-based method. Thus, it does not require a training phase, nor annotated data. We extract the line segments, in the image, at various resolutions. Then, a grammatical description combines those elements in a logical way to filter the relevant map contents. This is an original approach, at the era of deep learning techniques, but that is very convenient for old non-annotated documents. We validated this work during MapSeg competition. It obtained the third position for map segmentation and the second position for graduate line detection, with a localization score of 89.2%.
Fichier principal
Vignette du fichier
HIP_2021___Concours_map_segmentation.pdf (2.67 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03374571 , version 1 (12-10-2021)

Identifiants

Citer

Aurélie Lemaitre, Jean Camillerapp. Segmentation of historical maps without annotated data. 6th International Workshop on Historical Document Imaging and Processing (HIP’21), Sep 2021, Lausanne, France. pp.19-24, ⟨10.1145/3476887.3476909⟩. ⟨hal-03374571⟩
46 Consultations
185 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More