EDoF-NeRF uses coded apertures to extend depth-of-field in NeRF

Neural radiance fields (NeRF) have revolutionized 3D scene rendering, but they inherit a fundamental optical limitation: the trade-off between depth-of-field (DoF) and light sensitivity. When a camera captures images with a wide aperture, defocused regions lose high-frequency details, degrading the quality of NeRF reconstructions. To solve this, Yoshiyuki Shirasaki and Ryoichi Horisaki from Osaka University introduce EDoF-NeRF, which places a coded aperture—a specially patterned mask—at the camera pupil. Unlike a simple circular opening, a coded aperture diffracts light in a way that preserves spatial frequency information across a larger depth range. The team developed a camera model that directly feeds these coded images into the NeRF pipeline, allowing the neural representation to learn sharp features from defocused areas. In both simulation and real-world tests, EDoF-NeRF consistently outperformed standard aperture cameras, rendering novel views with significantly extended DoF.

The key insight is that the coded aperture acts as an optical pre-conditioner, encoding depth information into the blur pattern itself. This enables the neural network to deconvolve the defocus during training, resulting in high-fidelity 3D reconstructions without requiring extra hardware or light loss. The method also maintains compatibility with existing NeRF architectures—only the camera model is modified. By tackling the DoF-light tradeoff at the capture stage, EDoF-NeRF promises to make NeRF more practical for real-world applications where lighting is limited or scenes have large depth variations. The paper is published on arXiv and has implications for computer vision, robotics, and AR/VR content creation.