ICDAR 2021 Competition on Historical Map Segmentation

修订：v1.0.0-full-20210527a DOI：10.5281/zenodo.4817662 作者：J. Chazalon、E. Carlinet、Y. Chen、J. Perret、C. Mallet、B. Duménieu 和 T. Géraud 官方比赛网站：https ://icdar21-mapseg.github.io/ 这是 ICDAR 2021 历史地图分割竞赛（“MapSeg”）的数据集。本次比赛从 2020 年 11 月持续到 2021 年 4 月。 动机 本次比赛旨在鼓励对历史地图数字化的研究。为了可用于历史研究，需要提取此类图像中包含的信息。一般流水线涉及多个阶段；我们在这里列出一些基本的： 分段地图内容：定位包含地图内容的图像区域；从不同图层中提取地图对象：检测道路、建筑物、积木、河流等对象以创建几何数据；对地图进行地理配准：通过检测已知地理坐标处的对象，计算转换以将几何对象转换为地理对象（可以覆盖在当前地图上）。任务 我们提出的任务模拟了我们刚才提到的三个基本数字化步骤。任务一：“检测积木”这个任务是本次比赛的旗舰。给定一个专注于地图内容的地图片图像片段，您需要检测构建块。积木用粗线表示。它们之间不重叠，但许多其他元素会干扰它们的检测：特殊建筑物（阴影区域）可以包含在构建块中（有时它们完全覆盖构建块）；文本可以叠加在行上；这些地图包含许多需要过滤掉的线条（内部建筑结构、铁路、河流、花园……）。此任务的预期输出是一个二进制掩码，指示每个像素是否属于构建块。评估工具还允许使用 TIFF 格式的标签图，其中每个像素都使用 INT16 标记其所属形状的标识符。为了从二进制掩码中提取形状，使用了 4 连接（因此背景具有 8 连接）。任务2：“分割地图区域” 这个任务相当于OCR的文本区域检测：给定完整地图的图像，您需要分割包含地图内容的区域。该区域通常与其他元素（标题、图例、比例...）相隔几个框架，但有时地图内容超出了某些大型对象的框架。虽然大部分区域是用直线划定的，但一些物体是在几张纸上画在框架外的。我们决定尽可能地细分这些区域。该任务的预期输出是一个二进制掩码，指示每个像素是否属于地图区域。任务 3：“定位标线交叉点” 此任务对于地图的地理配准至关重要：标线是指示相对于参考点的北/南/东/西坐标的线。它们的交点对于为地图图像的配准提供关键点非常有用。给定完整地图的图像，您需要定位这些线的交点。这些线条通常从左到右或从上到下覆盖地图内容，但要注意：由于文件老化，纸张不再平整，线条不直；某些区域的线可能是对角线；线条可以与许多其他对象重叠。此任务的预期输出是坐标列表（在图像参考中，即左上角的 0,0，x 轴指向右侧，y 轴指向下方）。

Revision: v1.0.0-full-20210527a, DOI: 10.5281/zenodo.4817662, Authors: J. Chazalon, E. Carlinet, Y. Chen, J. Perret, C. Mallet, B. Duménieu and T. Géraud. Official competition website: https://icdar21-mapseg.github.io/
This is the dataset for the ICDAR 2021 Historical Map Segmentation Competition ("MapSeg"), which ran from November 2020 to April 2021.

### Motivation
This competition aims to promote research on historical map digitization. To enable utilization in historical research, information contained in such images must be extracted. The general workflow involves multiple stages; we list some fundamental ones below:
1. Map content segmentation: Locate image regions that contain map content;
2. Map object extraction from multiple layers: Detect objects such as roads, buildings, building blocks, rivers, etc., to create geometric data;
3. Map georegistration: Calculate transformations to convert geometric objects into geographic objects (which can be overlaid on current maps) by detecting objects at known geographic coordinates.

### Tasks
The tasks we propose simulate the three fundamental digitization steps mentioned earlier.

#### Task 1: "Block Detection"
This is the flagship task of this competition. Given an image patch of a map focusing on map content, you are required to detect building blocks. Building blocks are represented by thick lines. They do not overlap with each other, but many other elements interfere with their detection: Special buildings (shadowed areas) may be contained within building blocks (sometimes they completely cover the blocks); text may overlay the lines; these maps contain numerous lines that need to be filtered out (internal architectural structures, railways, rivers, gardens, etc.). The expected output for this task is a binary mask indicating whether each pixel belongs to a building block. The evaluation tool also allows the use of label maps in TIFF format, where each pixel is marked with the identifier of the shape it belongs to using INT16. To extract shapes from the binary mask, 4-connectivity is used (thus the background has 8-connectivity).

#### Task 2: "Map Region Segmentation"
This task is equivalent to text region detection in OCR: Given an image of a complete map, you need to segment the regions that contain map content. These regions are usually separated from other elements (titles, legends, scales, etc.) by frames, but sometimes map content extends beyond the frames of some large objects. While most regions are delineated by straight lines, some objects are drawn outside the frames across multiple sheets. We decided to subdivide these regions as precisely as possible. The expected output for this task is a binary mask indicating whether each pixel belongs to a map region.

#### Task 3: "Grid Line Intersection Localization"
This task is critical for map georegistration: Grid lines are lines that indicate north/south/east/west coordinates relative to a reference point. Their intersections are very useful for providing key points for map image registration. Given an image of a complete map, you need to locate the intersections of these lines. These lines usually cover the map content from left to right or top to bottom, but note: Due to paper aging, the paper is no longer flat, so the lines are not straight; lines in some areas may be diagonal; lines may overlap with many other objects. The expected output for this task is a list of coordinates (in the image reference frame, i.e., 0,0 at the top-left corner, with the x-axis pointing right and the y-axis pointing downwards).

搜集汇总

背景与挑战

背景概述

该数据集来自ICDAR 2021历史地图分割竞赛，旨在支持历史地图的数字化研究。它包含三个核心任务：检测积木、分割地图区域和定位标线交叉点，分别输出二进制掩码或坐标列表，以促进地图内容的提取和地理配准。