Abstract: Recent advances in camera technology have significantly improved image and video capture. However, challenging real-world conditions, including low-light environments, rapid object motion, and continuously changing illumination, can still severely degrade video quality and hinder scene understanding. This seminar presents three research contributions that address these challenges using videos captured with low-cost consumer cameras. Since capturing a scene from every viewpoint and at every instant is impractical, the proposed methods leverage a limited set of observations to reconstruct dynamic 3D scenes and synthesize novel views over time. The first contribution focuses on reconstructing sharp dynamic scenes from motion-blurred videos while simultaneously estimating the underlying object motion. The second addresses low-light video enhancement by recovering well-illuminated dynamic scenes, even when illumination varies spatially across different parts of the scene. The third tackles scenes captured under continuously changing coloured illumination, recovering their intrinsic appearance while separating static and dynamic scene components in 3D. These capabilities facilitate downstream applications such as scene editing and content creation. A key feature of all three methods is their self-supervised formulation. Furthermore, the latter two methods achieve real-time inference speeds exceeding 30 frames per second, demonstrating their potential for practical deployment. Collectively, these contributions advance the reconstruction and restoration of dynamic 3D scenes under challenging imaging conditions.

Event Details
Title: 3D Dynamic Scene Reconstruction and Novel View Synthesis under Diverse and Challenging real-world conditions
Date: October 23, 2026 at 12:00 PM
Venue: ESB 244 / Google Meet (https://meet.google.com/meet/mjh-yjug-dbj)
Speaker: Mr. Ashish Kumar (EE20D006)
Guide: Dr. Rajagopalan A N
Type: PHD seminar

Updated: