DGTR: Distributed Gaussian Turbo-Reconstruction for Sparse-View Vast Scenes

  • 2024-11-20 12:18:36
  • Hao Li, Yuanyuan Gao, Haosong Peng, Chenming Wu, Weicai Ye, Yufeng Zhan, Chen Zhao, Dingwen Zhang, Jingdong Wang, Junwei Han
  • 0

Abstract

Novel-view synthesis (NVS) approaches play a critical role in vast scenereconstruction. However, these methods rely heavily on dense image inputs andprolonged training times, making them unsuitable where computational resourcesare limited. Additionally, few-shot methods often struggle with poorreconstruction quality in vast environments. This paper presents DGTR, a noveldistributed framework for efficient Gaussian reconstruction for sparse-viewvast scenes. Our approach divides the scene into regions, processedindependently by drones with sparse image inputs. Using a feed-forward Gaussianmodel, we predict high-quality Gaussian primitives, followed by a globalalignment algorithm to ensure geometric consistency. Synthetic views and depthpriors are incorporated to further enhance training, while a distillation-basedmodel aggregation mechanism enables efficient reconstruction. Our methodachieves high-quality large-scale scene reconstruction and novel-view synthesisin significantly reduced training times, outperforming existing approaches inboth speed and scalability. We demonstrate the effectiveness of our frameworkon vast aerial scenes, achieving high-quality results within minutes. Code willreleased on our [https://3d-aigc.github.io/DGTR].

 

Quick Read (beta)

loading the full paper ...