Arch Recognition Dataset for Architectural Heritage Images
Description
This dataset supports the study on automatic arch recognition in architectural heritage images. The curated dataset comprises 7,479 images organized into two classes: Has Arch (4,384 images) and No Arch (3,095 images). The images were collected from two principal sources: Wikimedia Commons and the Ganjnameh architectural reference series. The study was designed to evaluate whether reliable automatic arch recognition can be achieved using existing and accessible architectural imagery collected from documentary and archival sources. A fixed stratified partition was used throughout the experiments, consisting of 5,235 training images, 748 validation images, and 1,496 test images. The deposited research files provide the dataset metadata, class information, and the exact fixed training, validation, and test partitions used in the reported experiments. The original third-party image files are not redistributed through this repository and remain subject to the reuse conditions of their respective sources.
Files
Steps to reproduce
The dataset was curated for binary image-level arch recognition in architectural heritage imagery. Images were manually organized into two classes: Has Arch and No Arch. The final dataset contains 7,479 images, including 4,384 Has Arch and 3,095 No Arch samples. A fixed stratified split with random seed 42 was used throughout the experiments: 5,235 images for training, 748 for validation, and 1,496 for testing. The same fixed test set was used for all reported model evaluations. Images were resized to 224 × 224 pixels and normalized using ImageNet mean and standard deviation values. The accompanying files provide the exact training, validation, and test membership used in the experiments. Source code and implementation details are provided separately in the associated code repository.
Institutions
- Shahid Beheshti UniversityTehran, Tehran