结合体素网格和神经辐射场的快速三维重建算法

三维重建作为计算机视觉的经典任务之一,在自动驾驶、增强现实、模拟现实等领域有着重要的应用,相较于传统的三维重建算法,神经辐射场具有重建精度高、适合复杂物体重建的优点;但神经辐射场的网络层数深、参数量大,导致训练时间长。为提高神经辐射场的训练速度,文章将神经网络与体素网格相结合,采用体素剪枝等经典算法减少采样点的数量,最终得到一个速度相较于原神经辐射场有极大提升的网络结构。实验结果表明,文章算法在DTU数据集上重建时间缩减至1 h左右,重建精度比原方法提高7%。在神经辐射场中使用体素剪枝和光线终止算法,不仅提高了训练速度,还提升了重建精度。

3D reconstruction, as one of the classic tasks in computer vision, has important applications in fields such as autonomous driving, augmented reality, and simulated reality. Compared to traditional 3D reconstruction algorithms, neural radiance fields have the advantages of high reconstruction accuracy and suitability for complex object reconstruction. However, neural radiance fields involve deep network layers and a large number of parameters, resulting in long training times. In order to improve the training speed of neural radiance field, this paper combines neural network with voxel grid, and reduces the number of sampling points by voxel pruning and other classical algorithms. Finally, a network structure with greatly improved speed compared to the original neural radiance field is obtained. Experiments show that the algorithm in this paper reduces the reconstruction time to about 1 hour on the DTU dataset and improves the reconstruction accuracy by 7% compared to the original method. Using voxel pruning and ray termination algorithms in neural radiance fields not only improves the training speed but also improves the reconstruction accuracy.