ExamHW-OCR (partⅠ)

Read full text Summarise with VIRAT
Authors: Zihan Deng
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

ExamHW-OCR Dataset Overview ExamHW-OCR is a large image dataset of real scanned answer-sheet regions collected from operational large-scale online marking / computer-based examination workflows in China. It covers scenarios commonly associated with senior-high entrance examinations (Zhongkao), college entrance examinations (Gaokao), mock exams, joint exams, and related standardized assessments. The release is organized as subject-specific subsets (course/paper codes) and is primarily intended for research on: Handwriting OCR Mathematical expression recognition Document layout understanding of examination papers Robustness to scanning artifacts Educational assessment analytics under realistic deployment conditions Data Organization & Structure The dataset contains 3,158,804 JPEG images (.jpg) in the examined backup snapshot. Images are stored in a hierarchical folder structure reflecting: Batch / session identifiers as top-level directories Examinee pseudonymous IDs (long numeric directory names) Cropped response regions – commonly 8 subfolders per examinee in the Current tree in sampled paths, consistent with multi-region cropping in marking pipelines A parallel tree, Current_jst, coexists with different per-batch image counts in sampled comparisons. This likely corresponds to an alternative processing stage or subset policy. Unless official documentation states otherwise, users should treat the two trees as distinct splits. Important Limitations & Notes No standalone annotation filesThe deposited snapshot does not include annotation files (e.g., COCO/VIA-style JSON/XML) alongside the image hierarchy. Ground truth (transcripts, LaTeX/MathML, question IDs, rubrics) may reside in external databases or proprietary export packages. Limited filename semanticsResearchers must not assume labels are embedded in filenames beyond internal operational encoding. Privacy and complianceBecause content originates from real student responses, any public redistribution requires privacy review, de-identification, and lawful basis compliant with applicable regulations and institutional policies.

Fields of Study & Subjects

Handwritten Text Recognition Techniques Educational Technology and Assessment Vehicle License Plate Recognition

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
21 March 2026
Language
CMN
Document type
Dataset