Explanaibility in Autonomous Driving

This was the starter project that Prof. Omranpour assigned me to get me started with deep learning and computer vision in general. It was an interesting project because I was not supposed to develop or come up with anything, and I was only supposed to learn and familiarize myself with as much as I could. I learned about a lot of ideas, techniques and research areas that I was not familiar with at that time. This project was, in my opinion, why I became so interested in computer vision and why I kept working in this field.

Face Recognition Loss (Collaborator)

In this project Arash (a member of Prof. Omranpour's group) is trying to develop a new loss function for face recognition. Loss functions are an important aspect of face recognition because of the large number of identities in the data and the security importance of correct identification. For this reason, we are trying to improve upon current works by introducing regularization terms based on FIQA factors.

Past-Attending Transformers

This project is my latest research in which I came up with the idea for it myself. When I was working on my bachelor's project on image restoration using transformers, I read many papers on different approaches to implementing transformers in vision tasks, none of which tried to explore any type of connection between QKV matrices of different layers (except for some memory-abled QKV models). This intrigued me to try to see further into this approach with the assistance of Prof. Omranpour. We have implemented this approach into two existing models in image restoration and image segmentation, and our approach has shown improve performance and convergence time. We are currently tweaking the base ideas and implementing some new ideas to improve our results even further.