LW-PointNeRF: A Point-Based Neural Radiance Field with a Lightweight Multi-View Stereo Network

作者：Muyuan Cheng, Zhengyao Bai, Haojie Chen, Guangming Wang, Yuee Xu, Shuai Song · 年份：2023 · DOI：10.1109/cac59555.2023.10451495 · 研究领域：CCD and CMOS Imaging Sensors、Functional Brain Connectivity Studies、Neural dynamics and brain function

In recent years, there have been significant advancements in novel view synthesis using Neural Radiance Fields ($N$eRF). NeRF's application is limited because of a considerable amount of training time. In this paper, we present a method called LW-PointNeRF. This method can rapidly reconstruct radiance fields while reducing the number of model parameters and computational requirements. Specifically, LW-PointNeRF comprises two main modules: the LW-MVS module and the Radiance Field module. The LW-MVS module employs a lightweight convolutional neural network to extract image features. It then uses plane sweeping to construct a cost volume for depth estimation, generating an initial point cloud. The Radiance Field module aligns the image features with the initial point cloud to get a neural point cloud. Finally, it aggregates the features of neighboring points to construct the radiance field. Experimental evaluations of LW-PointNeRF are conducted on the DTU, NeRF synthetic, and ScanNet datasets. LW-PointNeRF performs similarly to PointNeRF, but uses 25% fewer parameters and 14% less computation.