3D Gaussian Splatting (3DGS) has emerged as a powerful framework for scene reconstruction and novel view synthesis for normal‐light and illumination‐consistent scenes. However, the original 3DGS struggles to reconstruct scenes from low‐light RGB images, which are characterized by low pixel intensity and significant noise degradation. Near infrared (NIR) images captured under invisible near‐infrared light can maintain high signal‐to‐noise ratio and brightness in extremely dark environments but lack color information and suffer from interimage illumination inconsistency. In this article, we propose NIR‐assisted low‐light RGB Gaussian Splatting (NIRGB‐GS), a multimodal 3DGS‐based method for 3D reconstruction and normal‐light novel view synthesis of low‐light scenes. We adopt an NIR‐dominated Gaussian density control strategy to mitigate the influence of noise in RGB images and address the lighting inconsistency in NIR images with appearance encoding. Inverse histogram equalization is proposed for color and brightness recovery, even in extreme low‐light conditions. Experimental results across 7 real‐world low‐light scenes demonstrate that NIRGB‐GS is able to recover geometry, color, and texture of low‐light scenes and outperforms existing algorithms in both subjective perception and objective evaluation.
Yang et al. (Sun,) studied this question.