Recovering Scene Geometry and Material from Event Streams with 3D Gaussian Splatting

Recovering scene geometry and material yields physically meaningful representations that support many tasks in computer vision and graphics. Methods built on neural representations such as neural radiance fields and 3D Gaussian Splatting perform well but assume clean, well-exposed multi-view images, and their accuracy degrades substantially under low light, fast motion, or strong illumination changes. We instead recover scene geometry and material using only event streams (the asynchronous output of a neuromorphic event camera, which records per-pixel logarithmic-brightness changes above a contrast threshold) as input, within a framework based on 3D Gaussian Splatting. Because events measure differential brightness and leave the global brightness scale undetermined, we introduce a brightness-anchoring prior that constrains this global brightness scale while reconstructing a radiance field; we then decompose the radiance into geometry and material (surface normals, depth, base color, and roughness), using motion-compensated event edges to guide boundary-aware regularization. On public datasets, the method achieves the best average performance among baselines that first reconstruct images and then apply inverse rendering. Averaged over the seven evaluated scenes, it attains 22.11 dB PSNR, 0.902 SSIM, and 0.105 LPIPS on novel-view rendering, and 22.31 dB PSNR on base-color estimation, with a mean surface-normal angular error of 28.93∘. The quantitative evaluation uses simulated event streams on seven synthetic scenes; we additionally include a qualitative demonstration on a physical event-camera capture, and identify full quantitative validation on real recordings as the primary direction for future work.