ICDAR2017 Competition on Historical Document Writer Identification (Historical-WI)

Abstract

The ICDAR 2017 Competition on Historical Document Writer Identification is dedicated to record the most recent advances made in the field of writer identification. The goal of the writer identification task is the retrieval of pages, which have been written by the same author. The test dataset used in this competition consists of 3600 handwritten pages originating from 13 th to 20 th century. It contains manuscripts from 720 different writers where each writer contributed five pages. This paper describes the dataset, as well as the details of the competition. Five different institutions submitted six methods which were ranked using identification and retrieval metrics. The paper describes the competition details including the dataset, the evaluation measures used as well as a short description of each submitted method.

Deep Generalized Max Pooling

“…We evaluate the following publicly available datasets: 1) ICDAR17-WI: is a dataset composed of letters used in the ICDAR 2017 Writer Identification competition [31]. The training set contains 394 writers contributing three samples each.…”

ICDAR 2019 Competition on Image Retrieval for Historical Handwritten Documents

“…They come from the University Library Basel, 8 licensed under public domain. The images are non-used (and non-published) ones of the ICDAR'17 writer identification competition [1] but processed in a similar way. d) Charters: For a large part of the single images of the test set, we used charters from Monasterium.net.…”

“…Thereafter, they used the Fisher Vector [7] to encode the extracted SIFT features. The codebook was trained using the ICDAR17 historical writer identification dataset [1]. b) Pathlet: The second submitted method uses their newly proposed pathlet feature [8], which is aimed to capture useful information such as curvature and slant from the contours.…”

“…Pathlets were extracted from the binary images and encoded using VLAD encoding [9]. The codebook was also trained using the ICDAR17 dataset [1]. c) SIFT + Pathlet: The above two feature vectors were projected onto a subspace of lower dimensionality and then concatenated.…”

“…c) SIFT + Pathlet: The above two feature vectors were projected onto a subspace of lower dimensionality and then concatenated. The projection matrices were obtained by performing SVD on the feature matrices of the ICDAR17 dataset [1]. The concatenated vectors were further power-normalized and l2normalized and used as the final feature vectors.…”