MIDV-Copy

MIDV-Copy is a dataset designed for document presentation attacks detection task. It consists of three image classes: 1000 photographs of original documents, 1000 scans of original documents, and 2000 photocopies with a visible paper frame (1000 plain photocopies and 1000 photocopies with marks). Photographs and scans of original documents correspond to the "photo" and "scan_rotated" directories from the MIDV-2020 dataset (https://zenodo.org/records/18786808). Photocopies of documents are new data. Photocopies are produced with printing source scans in both grayscale and color. To obtain plain photocopies ("copy" directory), printed documents were captured with 5 mobile cameras. To obtain photocopies with marks ("copy_marks" directory), various stamps and notes were applied to the printed documents beforehand; the resulting printouts were also captured with 5 cameras. The markup contains document quadrilaterals, as well as paper quadrilaterals for photocopy images. The document image numbering is preserved across classes. The collection of images and their annotations from directories "photo", "scan_rotated", and "copy" constitutes the previously published dataset MIDV-Copy 2025 (ttps://doi.org/10.5281/zenodo.21242613). The current expanded version of the dataset, consisting of "photo", "scan_rotated", "copy" and "copy_marks" directories, constitutes dataset MIDV-Copy 2026. The base dataset (MIDV-Copy 2025) is published in: L. Tolstenko, A. Bursikov, I. Kunina "Detection of photocopied documents in remote identity verification by presence of the paper frame", Proc. SPIE 14114, Eighteenth International Conference on Machine Vision (ICMV 2025), 141142E (25 Feb 2026); https://doi.org/10.1117/12.3096443 The expanded version of the dataset refers to the work, which is currently in the publication stage. When the paper is published, the full reference will be added here.