Research Assistant to Dr. Fernando de la Torre, Associate Research Professor in Robotics Institute and director of Human Sensing Lab, and, Dr. Jessica K. Hodgins, aa Professor in the Robotics Institute and Computer Science Department at Carnegie Mellon University.

Research Overview

Professor Fernando de la Torre is a renowned expert in the field of Computer Vision and his research includes Augmented and Virtual Reality, Human Sensing and Data focused computer vision. We are currently working on leveraging synthetic data to satifiy the requirement of large training sets for deep learning models. Specifically, I am working on using synthetic data to train action recognition models and classifying actions in real domain.

My Contribution

1. Tested different versions of i3d on kinetics val data.
2. Modified the i3d non-local version to recognise and localise activity in a video.
3. Changed the input to the model, to fit Robot cognition data (synthetic and real) RoCoG.
4. Compared the accuracies by creating and training on different styles of synthetic data and tested on real domain.
5. Implemented gradient reversal layer for i3d to train the model on synthetic and real data without labels in real data to observe an increase in accuracy of 2%
6. Further, tuned the head of i3d to improve the f1 score of classes that are often being confused for.

Robot cognition data (RoCoG)