Xinran (Nicole) Han (Harvard): “Seeing Through Ambiguity: Generative Models for Perceptually-consistent Computer Vision”

Speaker: Xinran (Nicole) Han (Harvard)

Title: “Seeing Through Ambiguity: Generative Models for Perceptually-consistent Computer Vision”

Video: Click here to view video

Abstract:  Recovering 3D structure from 2D images is a central problem in computer vision, yet it is inherently ill-posed. In this talk, I argue that instead of seeking a single “best” estimate, computational vision systems should embrace this ambiguity by modeling the full distribution of plausible interpretations. I begin with static images of matte objects, demonstrating how generative models can produce diverse shape percepts that are consistent with ambiguities perceived by humans. Next, I expand the framework to arbitrary materials and incorporate object motion, presenting a method that leverages motion cues to disentangle shape and material information. Along the way, I highlight architectural choices that make this framework practical by integrating bottom-up and top-down processing. This line of work provides new ways to mimic human visual cognition and suggests new directions for building robust embodied systems.

Bio:  Xinran (Nicole) Han is a PhD candidate at Harvard University, advised by Prof. Todd Zickler. She received her bachelor’s degrees in Computer Science and Mathematics summa cum laude from the University of Pennsylvania. Her research interests span computer vision and human perception, with an emphasis on 3D understanding. Her work focuses on combining physics-based insights and learning-based neural priors to build data- and compute-efficient models that generalize to unseen scenarios.