Vesuvius challenge grand prize submission
We approached the ink detection task as a 3D-to-2D binary semantic segmentation problem using surface volumes from scroll 1 (PHerc Paris 3). We followed a human-assisted pseudo-label-based self-training approach using the crackle signal as a surrogate to the ink signal.
For a summary of the methods used, please see docs/methods.md.
For instructions on how to train and run inference, please see docs/submission_reproduction_instructions.md.
A pretrained checkpoint is available here (associated with val_3336_C3.yaml).
Louis Schlessinger, Arefeh Sherafati
- EduceLab-Scrolls: Verifiable Recovery of Text from Herculaneum Papyri using X-ray CT
- Introducing Hann windows for reducing edge-effects in patch-based image segmentation
- 1st place Kaggle Vesuvius Challenge - Ink Detection
- 4th place Kaggle Vesuvius Challenge - Ink Detection
- First Ink Vesuvius Challenge
- 2nd place Vesuvius Challenge First Letters