Fast and Generalizable NeRF Architecture Selection for Satellite Scene Reconstruction
This work addresses the problem of slow and scene-specific NeRF training for satellite scene reconstruction, offering a fast and generalizable solution that reduces computational costs and power usage on edge platforms.
The paper tackles the challenge of deploying Neural Radiance Fields (NeRF) for satellite imagery by developing PreSCAN, a predictive framework that estimates NeRF quality before training using lightweight descriptors, achieving a 1000× speedup over Neural Architecture Search with < 1 dB prediction error and enabling efficient architecture selection in < 30 seconds.
Neural Radiance Fields (NeRF) have emerged as a powerful approach for photorealistic 3D reconstruction from multi-view images. However, deploying NeRF for satellite imagery remains challenging. Each scene requires individual training, and optimizing architectures via Neural Architecture Search (NAS) demands hours to days of GPU time. While existing approaches focus on architectural improvements, our SHAP analysis reveals that multi-view consistency, rather than model architecture, determines reconstruction quality. Based on this insight, we develop PreSCAN, a predictive framework that estimates NeRF quality prior to training using lightweight geometric and photometric descriptors. PreSCAN selects suitable architectures in < 30 seconds with < 1 dB prediction error, achieving 1000$\times$ speedup over NAS. We further demonstrate PreSCAN's deployment utility on edge platforms (Jetson Orin), where combining its predictions with offline cost profiling reduces inference power by 26% and latency by 43% with minimal quality loss. Experiments on DFC2019 datasets confirm that PreSCAN generalizes across diverse satellite scenes without retraining.