Reproduce “ExploreGS: a vision-based low overhead framework for 3D scene reconstruction” on OpenResearch

About this paper

Yunji Feng, Chengpu Yu, Fengrui Ran, Zhi Yang, Yinni Liu

Researchers at Beijing Institute of Technology developed ExploreGS, a low-overhead, vision-based framework for drone-mounted edge devices that achieves photorealistic 3D scene reconstruction. The system processes scenes in 50 seconds on an NVIDIA Jetson Orin NX, providing complete and accurate models, contrasting with existing methods that often fail to reconstruct entire scenes.

Key insights

• Replacing traditional Structure from Motion (SfM) with a neural vision model (MASt3R) for camera pose estimation and geometry reconstruction significantly enhances robustness and efficiency, especially in challenging environments.
• An intelligent data filtering mechanism, the Match Selector, dramatically reduces redundant image processing by leveraging a Bag-of-Words model, cutting processed image pairs by 84-95%.
• Integrating autonomous exploration with on-device 3D Gaussian Splatting allows for complete, high-quality 3D reconstruction directly on resource-constrained edge hardware like the NVIDIA Jetson Orin NX.

Summary by alphaXiv.