【摘要】 为解决基于神经辐射场的三维重建存在对输入图像清晰度要求高、训练时间长的问题，提出面向三维重建的快速抗运动模糊神经辐射场方法。该方法以运动模糊图像成像原理为理论基础，通过高斯过程回归估计相机运动轨迹，并学习关键时间戳处的相机位姿；然后，联合训练高斯过程回归和体素表示的神经辐射场，从而拟合出更加准确的相机运动轨迹，快速实现抗运动模糊神经辐射场。该方法在多个运动模糊场景上验证有效性和快速性，其PSNR和SSIM分别达到27.86和0.708，在现有先进方法的基础上提升了约4.15%和10.45%,LPIPS降至0.305，下降了16.89%，并在保证三维重建性能的基础上，缩短了训练时间。更多还原

【Abstract】 This paper proposes a fast anti-motion-blur NeRF for three-dimensional(3D) reconstruction to address the problems of NeRF which has high requirements on the clarity of input images and long training time. The method is theoretically based on the motion blurred image imaging principle. The camera motion trajectory is estimated by Gaussian process regression and the camera poses at the key time stamp are learnt. The Gaussian process regression and the voxelbased NeRF are jointly trained. This results in a more accurate camera motion trajectory, which in turn achieves a fast antimotion-blur NeRF. The experimental results demonstrate that the proposed method achieved a PSNR of 27.86 and an SSIM of 0.708, which is an improvement of 4.15% and 10.45%, respectively, compared to existing advanced methods. The LPIPS decreased to 0.305, which is a decrease of 16.89%.更多还原