Frequency-based View Selection in Gaussian Splatting Reconstruction

Recent advances in photorealistic 3D reconstruction—ranging from Neural Radiance Fields (NeRF) to Gaussian Splatting—have demonstrated impressive results, but their heavy computational demands pose challenges for online robotics applications. In this paper, we address the critical problem of active view selection to minimize the number of input images required for high-quality Gaussian Splatting reconstructions as a robot traverses an environment. Our key insight lies in a novel frequency-domain ranking strategy that identifies the most informative viewpoints, enabling the system to capture essential scene features without redundant observations. Experimental results show that our method not only reduces computational overhead but also preserves high-fidelity reconstructions, making it a promising solution for efficient 3D modeling in robotics and other resource-constrained applications. The code is available at https://github.com/lemonci/gs-fft.