Semantic-Guided Point Cloud Upsampling Method for Visual Localization

Depth acquisition devices can be utilized in visual localization applications to construct an offline database comprised of images and point clouds, thereby facilitating more precise three-dimensional spatial positioning. Yet, the hardware limitations of such depth acquisition devices result in a significantly lower point cloud resolution when compared to the image pixel resolution. Furthermore, the point cloud inevitably contains a certain level of noise generated by environmental factors during the collection process. These factors considerably impede the precision of visual localization. Thus, this paper presents a semantic-guided method for point clouds upsampling. Leveraging the semantic distribution information of the image, this approach upsamples the point cloud data corresponding to pixels containing unique semantic information, providing denser point clouds while eliminating noise. Experimental results demonstrate the successful performance of the proposed approach in point cloud upsampling, thereby effectively enhancing the accuracy of visual localization.