ExploreGS: a vision-based low overhead framework for 3D scene reconstruction

TL;DR

This work tackles the high cost and overhead of lidar-based 3D mapping by proposing a vision-only framework for on-board drone reconstruction. ExploreGS integrates autonomous exploration, a BoW-based match selector, and a vision-model-based Gaussian Splatting pipeline to enable real-time, RGB-driven 3D reconstruction on edge devices, eliminating the need for pre-modeling flight paths or heavy SfM pipelines. Key contributions include a data-reduction strategy that cuts redundant imagery by 84–95% and an end-to-end autonomous pipeline that trains $3DGS$ models on-board. Across simulation and real-world tests, ExploreGS achieves approximately a $5\times$ speedup over baselines while preserving reconstruction quality, illustrating the practicality of low-cost, on-board aerial mapping and scene understanding for resource-constrained platforms.

Abstract

This paper proposes a low-overhead, vision-based 3D scene reconstruction framework for drones, named ExploreGS. By using RGB images, ExploreGS replaces traditional lidar-based point cloud acquisition process with a vision model, achieving a high-quality reconstruction at a lower cost. The framework integrates scene exploration and model reconstruction, and leverags a Bag-of-Words(BoW) model to enable real-time processing capabilities, therefore, the 3D Gaussian Splatting (3DGS) training can be executed on-board. Comprehensive experiments in both simulation and real-world environments demonstrate the efficiency and applicability of the ExploreGS framework on resource-constrained devices, while maintaining reconstruction quality comparable to state-of-the-art methods.

Figures (9)

• Figure 1: Traditional SfM algorithms can achieve good results in scenes with a single object (left), but they usually fail in scenes where the overlap between images in the sequence is small (right)
• Figure 2: Overview of our ExploreGS framework. We collect RGB data while exploring the unknown scene. Given two images with common viewing area as input, we get part point cloud continuously and combine them after all images are processed. With an initial point cloud generated, few GS trainings can achieve good result.
• Figure 3: The quadrotor of real-world test.
• Figure 4: Selection method of image acquisition in flight. The number of image pairs can be greatly reduced by comparing the similarity of common view area and screening the pictures with too large similarity.
• Figure 5: The gazebo world (a). The process of data collection and exploration in simulation (b).