Kevin Issac Raj

My research focuses on developing efficient and scalable methods for 3D computer vision, with an emphasis on reconstruction and generation in settings where data is sparse/noisy. While recent progress relies heavily on large multi-view models and dense supervision, these approaches often scale quadratically and are not efficient. I aim to address this by learning compact and structured representations of 3D scenes that reduce redundancy and improve efficiency. More broadly, I am also interested in how to incorporate geometric structure and local/global priors into generative models to improve multi-view consistency across viewpoints. Overall, during my PhD, I would like to build 3D vision systems that are data-efficient, computationally scalable, and capable of generalizing to real-world environments, ultimately supporting robust 3D scene reconstruction and generation.