{ "fetched_at": "2026-10-24T22:39:00Z", "source_url": "https://www.kaggle.com/competitions/cmhstm22t1/data", "pages": [ { "name": "data-description", "content": "## Dataset description\n\nThe participants of this competition will get the only available annotated dataset of Hebrew medieval manuscripts, VML\\_HP, which covers major Hebrew script types and modes. The training set contains 508 full manuscript pages picked from 125 manuscripts. The test set consists of 140 image pages picked from the manuscripts that are not present in the training set. Most of the manuscripts used in this dataset are kept in the British Library, the Bibliotheque Nationale de France, and the National Library of Israel. \n\n- **To ensure the fairness of the competition, we especially collected a _new_ test set of document images. This new test set is not part of the previous VML_HP releases.**\n\nTo simplify the classification task, the selected pages have only one script type and one script mode in the *main text* area. \nThe distribution of the script types and modes in the train set is presented in Table 1. \n\n![](https://i.imgur.com/yh0uB34.jpg)\n\nThe participants have to classify the script type of the *main* text area. The train.zip contains all the images for download.\nThe file \"train_hard.scv\" contains two columns: the first column - ImageID - contains the image name, and the second one is - SCRIPT TYPE - the classification to one of the 14 classes (as illustrated in Figure 1). \n\n![](https://imgur.com/fcjSiZw.png)\n\n**Figure 1: Ground truth example.**\n\n## Ground truth\nTo train their models, the participants are provided with a train set accompanied by the corresponding ground\ntruth file \"train_hard.scv\". The ground truth includes the image names and the corresponding hard labels (as illustrated in Figure 1). The labels correspond to the *main* text on each page and *ignore* marginal notes. The hard labels are defined according to the 14 classes, which are listed in Table 2.\n\n![](https://imgur.com/eEUFCMT.png)\n\n## Files\n\n* **train.zip** - the training set images\n* **train_hard.csv** - the training set labels\n* **test.zip** - the test set images (to be released on August 27, 2022)\n* **sample_submission.csv** - a sample submission file in the correct format\n\n## Columns\n\n* `ImageID` - image name\n* `SCRIPT TYPE` - is a label (one of 14 classes listed in Table 2)", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Description", "content": "This competition is organized in conjunction with the International Conference of Frontiers in Handwriting Recognition [ICFHR2022](http://icfhr2022.org/), which will be held in Hyderabad, India, on December 4-7.\n\n### The competition consists of two tasks: (1) hard classification and (2) soft classification. Each participant can participate in both or only one of them. Here, you are competing in TASK 1. To register and compete in [TASK 2](https://www.kaggle.com/competitions/cmhstm22t2), follow this[ link](https://www.kaggle.com/t/3453f441f1f44bb1b9b8008ba832c839).\n\nAutomatic script classification has been an active area of research in recent years. It is essential for understanding, authenticating, and dating historical texts. Script classification is a prerequisite for many historical document processing tasks, such as paleography analysis, writer identification, and handwritten text recognition (HTR).\n\nThis competition adds the challenge of classification of medieval Hebrew scripts based on their type and mode.\n\nThe task to be evaluated in the present competition is classifying image pages written in Hebrew script. There are 508 pages in the training set that were picked from 125 manuscripts dating from the 10th to 16th centuries. The results will be evaluated on 140 image pages picked from manuscripts that are not included in the training dataset.\n\nIn the hard classification task (TASK 1), there is one label for each page image in the data set.\n\n\nThe organizers provide a training data set together with their ground truth in the form of hard labels. \n\n\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Evaluation", "content": "\n### Evaluation \nThe evaluation of Task 1 (hard classification) is based on average accuracy. [Accuracy](https://en.wikipedia.org/wiki/Accuracy_and_precision) is the fraction of correct classifications:\n\n$$ Accuracy = \\frac{correct\\ classifications}{all\\ classifications}$$\n\nIn addition, each participant is required to provide a short description of the method.\n\nTo check the applicability of the proposed systems for real-time application, applicants are also requested to report the processing time per document page. \n\n### Input and output format specifications\n\nThe input to the classifier is a document page image. The output is a single value (class) for TASK 1.\n\n### Submission Format\n**For every image in the test set**, the submission file should contain two columns: `ImageID` and `SCRIPT TYPE,` as illustrated in the \"sample_submission.csv\" file. The file should contain a header and have the following format:\n![](https://imgur.com/fcjSiZw.png)\n\n**Figure 1: Submission format.**\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "rules", "content": "### Participation\n1. Each participant can participate in both or only one of the classification tasks\n2. An article about this contest will be published in the [ICFHR2022](http://icfhr2022.org/) proceedings. By making a submission, the participants will be asked to provide a short description of the method.\n3. Winning participants are required to provide code as well as a description of their method. The competition hosts need to replicate the results before awarding any prize money. \n\n\n### One account per participant\nParticipants cannot sign up to Kaggle from multiple accounts and, therefore, cannot submit from multiple accounts.\n\n### No private sharing outside teams\n\nPrivately sharing code or data outside of teams is not permitted. It's okay to share code if made available to all participants on the forums.\n\n### Team Limits\nThere is no maximum team size.\n\n### Submission Limits\nParticipants may submit a maximum of 2 entries per day.\n\n### The use of the competition data is not permitted for commercial purposes.\nThe use of the competition data is permitted for research purposes with the condition of properly citing the article mentioned above in any related publication.\n\n### Only automatic prediction methods are allowed to compete. \n\n## Prizes\nThe first- and second-place winners of this competition will receive a prize of $300 and $200, respectively, at a special session at the [ICFHR2022](http://icfhr2022.org/) conference.\n\n\n\n\n\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Prizes", "content": "## Prizes\nThe first- and second-place winners of this competition will receive a prize of $300 and $200, respectively, at a special session at the [ICFHR2022](http://icfhr2022.org/) conference.\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Hebrew Script Types And Modes", "content": "#Hebrew Script Types And Modes\n\nHebrew manuscripts refer to manuscripts written in Hebrew characters, as the language was often adopted from\nthe host societies (Ladino, Judeo-Arabic, Yiddish, etc.). Geographically, the spread of the Hebrew manuscripts\nwas larger than Latin, Greek, or Arabic manuscripts. Hebrew scripts were influenced by the local traditions\nand often resembled the manuscripts of the host societies in scribal manner, material, and ways of production.\nThe survival rate of Hebrew manuscripts is much lower than that of Latin, Arabic, or Greek ones.\nThere are 14 major classes of Hebrew script defined by the type and the mode. These classes are based on\nthe definition of the types and modes formulated in the works of Malachi Beit-Arié, Norman Golb, Benjamin\nRichler, Colette Sirat [28, 29, 30, 31, 32, 33, 34] and are summarized in Table 2.\nThere are six main types of the Hebrew script: Oriental, Sefardic, Ashkenazi, Italian, Byzantine, and\nYemenite. Each script type may contain up to three graphical modes: square, semi-square, and cursive. Figure 2\npresents sample patches for each class. Semi-square (semi-cursive) and cursive modes in each regional style are\ndefined relative to the style’s square mode. A graphical mode in one regional style can look similar to a different\nmode in another geographical style, e.g., the square mode in Italian can be similar to the Ashkenazi semi-square\nmode. Styles and modes of Hebrew scripts, paleographically, can not be separated.\n![](https://imgur.com/eEUFCMT.png)\n\n![](https://imgur.com/BT4pqhS.png)\n\n**Figure 2: Examples of the documents from each of the 14 defined classes**\n\n\n### References\n\n[28] M. Beit-Arié, Hebrew Codicology. ZFDM Repository, 2021.\n[29] M. Beit-Arié, “Hebrew codicology,” Tentative Typology of Technical Practices Employed in Hebrew Dated\nMedieval Manuscripts. Jerusalem, 1981.\n[30] M. Beit-Arié and E. Engel, Specimens of medieval Hebrew scripts, in 3 vol. Israel Academy of Sciences\nand Humanities, 1987, 2002, 2017.\n[31] C. Sirat, Hebrew manuscripts of the Middle Ages. Cambridge University Press, 2002.\n[32] A. Yardeni et al., The book of Hebrew script: history, palaeography, script styles, calligraphy & design.\nCarta Jerusalem, 1997.\n[33] B. Richler and M. Beit-Arié, “Hebrew manuscripts in the Biblioteca palatina in parma: catalogue; palaeo-\ngraphical and codicological descriptions,” 2011.\n[34] B. Richler, “Hebrew manuscripts in the vatican library: catalogue,” Hebrew manuscripts in the Vatican\nLibrary, pp. 1–790, 2008.\n\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Competition Timeline", "content": "### Competition Timeline\n\n- July 17, 2022: Competitions are published on the website and opened to participants.\n- July 17, 2022: Training dataset available for registered participants.\n- August 27, 2022: Test dataset available.\n- September 1, 2022, 23:59 UTC: Delivery of results with a two-page description of the method.\n- September 15, 2022,23:59 UTC: Submission of competition reports.\n- September 25, 2022,23:59 UTC: Submission of full paper review describing the competition.", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Related Topics", "content": "\nScript classification is essential for understanding, authenticating, and dating historical texts. It is a prerequisite\nfor many historical document processing tasks, such as paleography analysis, writer identification, and text\nrecognition (HTR). Identifying the script type can aid text recognition of medieval manuscripts. The same letter\ncan be written differently in different script types and modes, and there are sometimes significant variations\nwithin a given type. In addition, the Hebrew alphabet was used for languages other than Hebrew, for example,\nJudeo-Arabic and Ladino.\nDue to the historical continuity in the evolution of script types along the temporal axis and the geographic\nregions, a script may incorporate characteristics from multiple classes. Human experts tend to classify script\ntype in a subjective manner, which is often influenced by their training school or background. In contrast,\nautomatic script categorization can provide historical research with the sometimes absent quality of objectivity\nand the benefit of objective verification of results.\nThe present competition is related to the following images analysis tasks:\n- Writer identification, verification, and retrieval\n- Paleography analysis\n- Word spotting and handwritten text recognition\n- Document layout analysis and segmentation", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Organizing Team", "content": "The competition is organized by:\n- **Dr. Irina Rabaev** (irinar@ac.sce.ac.il),\nSoftware Engineering Department,\nShamoon College of Engineering, Beer Sheva,\n84100, Israel\n- **Dr. Daria Vasyutinsky** (dariavas@post.bgu.ac.il),\nComputer Science Department,\nBen-Gurion University of the Negev, Beer Sheva,\nIsrael\n- **Mr. Ahmad Droby** (drobya@post.bgu.ac.il),\nComputer Science Department,\nBen-Gurion University of the Negev, Beer Sheva,\nIsrael\n- **Ms. Reem Alasam** (rym@post.bgu.ac.il),\nComputer Science Department,\nBen-Gurion University of the Negev, Beer Sheva,\nIsrael\n- **Prof. Jihad El-Sana**(el-sana@cs.bgu.ac.il),\nComputer Science Department,\nBen-Gurion University of the Negev, Beer Sheva,\nIsrael", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false } ] }