The Super AI Engineer 2021: Handwritten Digit competition requires particpant to predict 1-5 digits from the images pool captured by participants themselves. We applied various technique and ultimately select TrOCR to be used in this notebook.
Unfortunately, the competition dataset is not available publicly. To fully use the notebook, you have to be competition's participant and change Google Drive path in the notebook to your own images storage path.
- Ranked #3 amongs 50 teams participating the Kaggle hackathon in the score leaderboard.
- Scored (Edit distance) 0.31059 and 0.30506 in private and public leaderboard
Because the competition has submission limit, we manually labelled the 870 samples competition datasets for testing purpose only. The labels are provided in this repository as labels.csv
- Norapat Buppodom: Modeling and Programming
LinkedIn
GitHub
- Teeraphong Kanchanakantikul: Modeling and Visualization ๐ Kaggle
https://www.kaggle.com/kengmut
- Super AI Engineer program for creating this program/competition and kindly support the development of AI Engineers in Thailand.
- Khun Thitorn: ORAND Dataset and TrOCR inspiration
LinkedIn
- Huggingface tutorial and TrOCR Original paper
- ARDIS: A Swedish Handwritten Digit Dataset