Subtasks | Datasets | Models | Contributors
In this repo, we provide our solution to solve two subtasks of International Competition on Few-Shot and Many-Shot Layout Segmentation of Ancient Manuscripts.
-
Task 1: Few-Shot Layout Segmentation. Create an effective document layout segmentation system using only three images for each manuscript for training and ten additional images, with the corresponding ground truth, are provided for validation only.
-
Task 2: Many-Shot Layout Segmentation. Create a layout segmentation system using 35 images per manuscript along with their corresponding ground truth, divided into training, validation, and test sets.
The datasets can be accessed at here.
We follow the proposed architecture in Rahai et al. (2023), which is L-U-Net-based architecture with a three-step procedure. The details of the experiment setups are shown in the report.