ATM-NeRF: Learning Adaptive Tone Mapping for Normal-Light Neural Radiance Field Reconstruction

NeRF has gained significant attention for its effectiveness in novel view synthesis. However, the ideal-exposure assumption on input images does not account for lighting degradation such as low-light or overexposure, which leads to suboptimal performance in adverse lighting conditions. Low-light images suffer from noise and poor visibility, while overexposed images lose details due to highlight clipping. To address these issues, we propose a novel method, ATM-NeRF, that reconstructs high-quality and exposure-corrected radiance fields from multi-view low-light or overexposed images within a unified framework. Specifically, we leverage a learnable tone mapping function to adaptively adjust the scene's brightness and contrast according to exposure conditions, enabling the recovery of visually appealing results. Experimental results on the LOM dataset show the superior performance of ATM-NeRF in rendering well-exposed images compared with NeRF-based and 2D enhancement methods. Our code will be publicly available.