My earliest introduction to CV, like any other aspirant came through learning a dog-cat classification using a vanilla CNN. Over the course of time, I was introduced to more complex architectures. What attracted me was how closely we use fundamentals of mathematics (a favourite example being WGAN and Kantorovich-Rubenstein Duality) in various CV&PR tasks. A current problem that I would like to solve in this area is redefining the requirement of image-to-image models especially in the field of AVs to tackle the Sim2Real gap. But, my studies have putforth a question to me: "Does the Sim2Real gap exist?"
You can consider me as one of those "weirdos" who suprisingly loves mathematics and I can strongly state the world itself corresponds to mathematics and nothing else. Over the course of time, amidst reading a good amount of papers, I was drawn into the exploration of the true mathematics behind various deep neural nets. Though I have a long way to go and do not have a published work in this area, I do feel this is one area that requires a lot of answers in ML/DL/AI. What I like looking into here includes learning how to prove various ML theorems, involvement of math in CV and probability and uncertainty.
This is one of the most interesting problems to ever exist if you ask me! I have a huge liking towards the concept of autonomous agents especially autonomous vision and have the habit of constantly following up with a few trends in this field. I really would love exploring the idea of building intelligence for AV, and especially looking into the capabilities of simulators like CARLA in the current era. Another interesting topic that I would love to explore in this area is quantifying uncertainty in self driving cars.