Short Bio:
Madhu Vankadari is a doctoral candidate at the University of Oxford’s Cyber Physical Systems group, under the supervision of Prof. Niki Trigoni and Prof. Andrew Markham. Prior to Oxford, he worked as a Machine Vision researcher at TCS Research in India. Madhu’s research revolves around using deep learning for SLAM-related challenges, such as improving depth estimation, camera pose accuracy, multi-motion scenarios, and visual place recognition. His work finds applications in robotics and computer vision, enhancing areas like autonomous navigation and augmented reality.
Abstract:
Understanding the world in 3D irrespective of the time of the day is crucial for applications such as autonomous navigation, and augmented and virtual reality. Amongst all the sensors through which this can be achieved, cameras have been cheap and ubiquitous. However, cameras can only capture the 2D projection of the 3D world. Extracting 3D information from one or more 2D images has been a long-standing problem in Computer Vision. Recently, the success of deep learning has made it possible to do the aforementioned by training a network on a large corpus of training data with their ground truth. Self-supervised learning made it possible to train a system to achieve the same objective without using any ground truth. In this talk, I am going to present some of the latest advances in self-supervised learning including my own research in this direction.
