Abstract: Radiance Fields, including Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS), have revolutionized novel view synthesis and 3D scene reconstruction. However, they inherently operate under the strict “Ideal Capture Assumption,” failing catastrophically when faced with real-world physical artifacts like complex optical lens flares and extreme motion blur. To address this bottleneck, this thesis proposes a unified framework that integrates computational imaging strategies with Radiance Fields to successfully navigate these severe spatio-temporal degradations.

The core contributions of this thesis systematically address both spatial-optical and temporal imaging constraints. To mitigate optical degradations, we introduce GN-FR, a generalizable NeRF framework that frames flare removal as a multi-view 3D reconstruction problem rather than a 2D filtering task. We then transition this robustification to the explicit, real-time 3DGS domain via FLAVORS, utilizing a novel Flare-Robust Exposure Alignment Loss (FREAL) to explicitly disentangle transient optical aberrations from stable scene geometry. To tackle extreme temporal degradations, we present PRISM3D, an efficient framework that recovers scenes from severely motion-blurred images by coupling deep dense tracking with probabilistic densification and continuous Bézier trajectories. Pushing the inverse problem to its theoretical limit, we finally introduce BeSplat, which leverages asynchronous event streams to successfully recover a sharp 3D radiance field and continuous camera ego-motion directly from a single severely motion-blurred image.

Collectively, these contributions establish a highly resilient framework for spatio-temporal perception and synthesis. By seamlessly integrating computational imaging modalities with physical degradation models, this work effectively breaks the reliance on the ideal capture assumption. These advancements push the capabilities of radiance fields far beyond the constraints of pristine captures, paving the way for reliable, high-fidelity deployment in complex, unconstrained real-world environments.

Event Details
Title: Resilient Radiance Fields: Robust 3D Scene Reconstruction Under Severe Optical and Temporal Degradations(PhD viva voce)
Date: July 01, 2026 at 03:00 PM
Venue: Google Meet (https://meet.google.com/can-bssx-aye)
Speaker: Mr. Matta Gopi Raju (EE17D021)
Guide: Dr. Kaushik Mitra
Type: PHD seminar

Updated: