Heimatkunde dataset

Multi-modal layout analysis and classification dataset.

Experiments

Experiments performed on this dataset are hosted on a different GitHub repository, which can be accessed here: https://github.com/honzikv/multimodal-document-processing-thesis

Structure

The root of the repository constitutes the root of the dataset. The structure is as follows:

• images - default folder for COCO images
• test.json and train.json are the COCO annotations for the test and train splits respectively. These files are usable e.g. in Detectron2 for instance segmentation or object detection.
• classifier - contains the classification dataset, which are mapped annotations to our custom format that is loadable with scripts from our experiments
• yolo - contains the YOLO annotations for the dataset which are also applicable to object detection or instance segmentation
• ocr - contains the OCR annotations for the dataset