NeRF-based Single Object Material Property Estimation for Lighting Simulation

This thesis investigates the problem of inserting relightable scanned objects into Neural Radiance Field (NeRF)-based scene reconstructions without retraining the scene or relying on ray-traced rendering. Traditional neural rendering pipelines offer high visual fidelity but lack explicit control over geometry and lighting, making object insertion a complex challenge. To address this, a modular pipeline is proposed that combines TorchNGP for efficient scene reconstruction with NeRFactor for estimating material properties and relighting scanned objects. The approach enables realistic object appearance by estimating a local environment map from the NeRF scene and using it to relight the object based on its learned BRDF parameters. The insertion is performed at render time through a custom ray marching strategy that blends the object and scene colors based on bounding-box intersections. This avoids modifying or retraining the original NeRF and allows for object reuse across multiple scenes. Evaluation is conducted using a custom Blender-based setup that provides ground truth renderings for photometric and perceptual comparison. While the system successfully demonstrates plausible object relighting and occlusion handling, it lacks support for mutual illumination, shadows, and reflections, which limits overall realism. Challenges with material decomposition and scene integration further highlight the trade-offs of lightweight, modular solutions.