Three-dimensional reconstruction is a key technology employed to represent virtual reality in the real world, which is valuable in computer vision. Large-scale 3D models have broad application prospects in the fields of smart cities, navigation, virtual tourism, disaster warning, and search-and-rescue missions. Unfortunately, most image-based studies currently prioritize the speed and accuracy of 3D reconstruction in indoor scenes. While there are some studies that address large-scale scenes, there has been a lack of systematic comprehensive efforts to bring together the advancements made in the field of 3D reconstruction in large-scale scenes. Hence, this paper presents a comprehensive overview of a 3D reconstruction technique that utilizes multi-view imagery from large-scale scenes. In this article, a comprehensive summary and analysis of vision-based 3D reconstruction technology for large-scale scenes are presented. The 3D reconstruction algorithms are extensively categorized into traditional and learning-based methods. Furthermore, these methods can be categorized based on whether the sensor actively illuminates objects with light sources, resulting in two categories: active and passive methods. Two active methods, namely, structured light and laser scanning, are briefly introduced. The focus then shifts to structure from motion (SfM), stereo matching, and multi-view stereo (MVS), encompassing both traditional and learning-based approaches. Additionally, a novel approach of neural-radiance-field-based 3D reconstruction is introduced. The workflow and improvements in large-scale scenes are elaborated upon. Subsequently, some well-known datasets and evaluation metrics for various 3D reconstruction tasks are introduced. Lastly, a summary of the challenges encountered in the application of 3D reconstruction technology in large-scale outdoor scenes is provided, along with predictions for future trends in development.

The emergence of Gaussian Splatting has ushered in a new phase in large-scale scene geometry reconstruction. Existing research in this field primarily focuses on computational efficiency and often neglects the effective utilization of inherent structural geometric priors. This paper introduces a shape-optimized Gaussian Splatting method based on the Manhattan World Assumption for precise geometric reconstruction from drone imagery. By leveraging the constraints provided by this assumption, a geometry-driven optimization mechanism is developed to guide the deformation and distribution of each Gaussian splat. This ensures improved alignment with the dominant structural directions of the scene. Additionally, dense point clouds are generated from the optimized Gaussian representation, making them more suitable for downstream tasks. The proposed approach maintains high visual fidelity while significantly improving geometric accuracy. Extensive experiments on multiple large-scale scene datasets, including a new benchmark for Jinan City, demonstrate that the method surpasses current state-of-the-art approaches in geometry-oriented evaluation metrics by a considerable margin.

3D Gaussian Splatting (3DGS) has significantly advanced the field of visual SLAM. However, outdoor monocular SLAM systems still suffer from scale ambiguity, so recovering the absolute scale remains difficult, and tracking often becomes brittle in large scenes. To address these issues, we present MoGe2-SLAM, an outdoor monocular 3DGS-SLAM framework that incrementally recovers camera trajectories and dense maps in an approximate absolute metric scale from a monocular RGB sequence with a known camera-height prior. By approximate absolute metric scale, we mean that camera poses and the reconstructed map are expressed in meters after camera-height calibration, rather than only up to an unknown global scale. Our core idea is to anchor frozen MoGe-2 geometry to a physical camera-height/ground-plane ruler, enabling approximate absolute-scale reconstruction. Specifically, we perform Absolute-Scale Initialization on the first keyframe by replacing MoGe-2’s advertised metric global scale with a known camera height and ground-plane geometry, generating the initial Gaussian scene directly in physical units. We then adopt Grid-Balanced Tracking that encourages spatially distributed feature extraction and maintains stable tracking. For mapping, we employ an ORB-anchored strategy that aligns the MoGe-2 depth prior with the rendered depth using geometric constraints, thereby mitigating scale drift. On KITTI, Waymo, and Argoverse 2, MoGe2-SLAM is competitive with existing monocular SLAM baselines in trajectory shape and novel-view fidelity while recovering trajectories and Gaussian maps in an approximate physical scale under the stated camera-height prior.