SuperNeRF: High-Precision 3-D Reconstruction for Large-Scale Scenes

Recent approaches based on neural radiance field (NeRF) showcase remarkable results in the 3-D reconstruction of small-scale scenes by encoding volume density and color observations using implicit functions. However, when confronted with complex and diverse large-scale scenes, it invariably experiences issues such as blurry textures and missing details. In this work, we present a superpixel-based neural radiance field (named SuperNeRF), an additional loss for learning radiance fields that takes advantage of superpixels texture constraints. Building upon NeRF, we leverage superpixels to establish spatial consistency constraints, enabling the precise extraction of 3-D geometry and appearance for large-scale scenes. SuperNeRF is capable of guiding locally adjacent and similar pixels to form nearly consistent ray termination distributions, and it is compatible with the state-of-the-art NeRF-based methods. Comprehensive experiments conducted on representative aviation and aerospace datasets demonstrate that our SuperNeRF exhibits a significant superiority in accuracy over state-of-the-art methods. Code will be available at https://github.com/xczbecalm/supernerf.

I. Introduction

Large-scale 3-D reconstruction based on 2-D images is a core task in the field of photogrammetry and remote sensing (PRS). Reconstructed 3-D structure has widely applied in city management [1], urban and rural planning [2], heritage protection [3], building damage assessment [4], estimation of the potential achievable solar energy of the buildings [5], and so forth. In PRS, the multiple-view geometry method [6], [7], [8], [9], [10], [11] is one of the main methods for the 3-D reconstruction of large-scale outdoor scenes. The pipeline of 3-D reconstruction based on multiple view geometry methods [12], [13], [14] can generally be divided into the following steps. 1)

Feature Extraction and Matching: Extracting distinctive features from multiple images and matching them across views.

2)

Camera Calibration: Estimating intrinsic and extrinsic parameters of the cameras used to capture the images.

3)

Structure From Motion: Estimating the 3-D camera poses and sparse point cloud of the scene.

4)

Dense Reconstruction: Creating a dense 3-D model by triangulating pixel correspondences from multiple views.

5)

Surface Reconstruction: Creating a surface mesh from the dense point cloud to represent the 3-D object.

6)

Texture Mapping: Mapping texture information extracted from images onto the reconstructed 3-D model to increase realism and detail.