Abstract
<title>Abstract</title> <p>3D Gaussian Splatting (3DGS) has emerged as a landmark paradigm for real-time radiance field rendering and high-fidelity scene reconstruction. However, under severe low-light conditions, its reliance on photometric consistency fundamentally breaks down due to a low signal-to-noise ratio (SNR). Noisy photometric gradients mislead the densification mechanism into over-proliferating Gaussian primitives, producing floating artifacts in empty space and corrupting the underlying 3D geometry. To address these challenges, we introduce EDD-3DGS, an explicit depth-distillation framework tailored for low-light 3D Gaussian Splatting. Our framework incorporates an Illumination-Inverted Loss Balancing scheme that dynamically weights photometric and depth prior supervision according to local sRGB luminance, effectively deactivating noise-fitting photometric loss in dark regions while enforcing geometric regularization. To cope with sparse and noisy monocular depth priors in under-exposed areas, we propose a Sparse-Aware Confidence Gate (SC-Gate) paired with a Weber-contrast Bilateral Total Variation (TV) loss, which preserves high-frequency structural boundaries while promoting piecewise-smooth depth surfaces. Furthermore, we devise a per-view Closed-Loop Depth Distillation strategy that continuously aligns the scale and shift of monocular depth priors using the rendered 3DGS geometry, entirely avoiding costly Vision Transformer fine-tuning. Extensive quantitative and qualitative experiments on low-light benchmarks demonstrate that EDD-3DGS effectively eliminates floating artifacts, retains sharp geometric boundaries, and achieves a 16.49% reduction in the number of primitives over vanilla 3DGS, all while maintaining state-of-the-art rendering quality.</p>