Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Continual learning for NeRF in scenes obtained from drones

Angelidis Nikolaos

Full record

---

Summary

In recent years, climate change has increased the frequency and severity of natural disasters such as earthquakes, floods, and wildfires. In the immediate aftermath of such events, fast and accurate 3D reconstructions of affected areas can support critical decision making for emergency response and recovery. Neural Radiance Fields (NeRFs) have emerged as a powerful solution for novel-view synthesis and 3D reconstruction from sparse imagery. However, traditional NeRF pipelines assume static scenes and require full retraining whenever updates occur, making them unsuitable for dynamic, real-world environments. In this thesis, we explore continual learning approaches to extend NeRF models with the ability to incrementally update scene reconstructions without catastrophic forgetting. We evaluate two frameworks, Nerfstudio and CLNeRF, where custom tooling is developed to support continual learning scenarios. While Nerfstudio showed promise, technical limitations led us to focus primarily on CLNeRF, which we extended and adapted to suit a variety of experimental conditions. We assess performance across a diverse set of scene changes, including object additions/removals, lighting variations, occlusions, and with a variety of types of input datasets. Our findings demonstrate that continual learning methods can significantly reduce training time, maintain reconstruction quality, and handle complex scene dynamics. These results lay the groundwork for deploying systems using NeRFs in time-sensitive scenarios such as mapping of a post-disaster scene using drone imagery.

Available Files

Services