An attempt at the Quick, Draw! Doodle Recognition Kaggle Challenge. We present an ensemble of a BiLSTM and a ResNet-18, which achieves a MAP@3 score of 0.917 on the Kaggle leaderboard on training with just 10% of the data.
This project was undertaken as a part of the Fall-2018 Introduction to Deep Learning course offered by the School of Computer Science, Carnegie Mellon University. The preprocessed dataset can be found here and the trained model can be found here. These files should replace the placeholders. A paper describing the architecture can be found here
- Tanmaya Shekhar Dabral
- Ramesh Balaji
- Ernest Yucheng Chang
- Irina Javed