NeRF-VLD: Efficient Visual Landmark Database Construction via Scene Constraints

Abstract

Visual landmarks, with their unique textures, play a crucial role in vision-based environmental sensing applications (such as augmented reality-based advertising, indoor navigation). However, few studies have focused on the construction of visual landmark databases, which is usually time-consuming and laborious. We propose a novel method called NeRF-VLD, which first introduces Neural Radiance Fields (NeRF) to build visual landmark databases efficiently. NeRF-VLD generates images from a small number of sources, thus eliminating the need for extensive site surveys and consequently reducing deployment costs significantly. To improve the quality of generated images, we introduce two scene-constrained modules. In the first module, sparse depth maps provided by Structure-from-Motion are introduced as additional geometric constraints. In the second module, scene regularization is applied to image pixels where depth information is unavailable, constraining the volume density of sample points near the camera plane to alleviate ghostly artifacts problem. Experimental results demonstrate that NeRF-VLD can achieve comparable landmark identification and positioning accuracy to real-image databases while reducing the number of image acquisition by 97%. Compared with full manual survey, NeRF-VLD reduces survey and update costs and improves the applicability of vision-based service.