CULTURE3D: A Large-Scale and Diverse Dataset of Cultural Landmarks and Terrains for Gaussian-Based Scene Rendering

TL;DR

CULTURE3D tackles the bottleneck of large-scale, high-fidelity 3D reconstruction by introducing a public dataset of $41{,}006$ 48MP drone images spanning $20$ culturally significant sites, totaling $10^{10}$ points. The work provides raw imagery, 3D models, and a full data-processing pipeline to enable dense reconstructions and textured assets, with COLMAP-friendly outputs to support broad benchmarking. It systematically benchmarks contemporary Gaussian-based rendering methods (e.g., 3D Gaussian Splatting, SuGaR, GOF, Wild Gaussian, City Gaussian) on CULTURE3D, revealing memory constraints and failure modes while highlighting City Gaussian as a robust baseline for large-scale heritage scenes. The study demonstrates CULTURE3D's value as a rigorous, diverse benchmark that drives memory-efficient, high-fidelity reconstruction research applicable to cultural heritage, urban-scale visualization, and VR/navigable environments, and it calls for future work on scalability, dynamic-object handling, and hybrid neural-geometric approaches.

Abstract

Current state-of-the-art 3D reconstruction models face limitations in building extra-large scale outdoor scenes, primarily due to the lack of sufficiently large-scale and detailed datasets. In this paper, we present a extra-large fine-grained dataset with 10 billion points composed of 41,006 drone-captured high-resolution aerial images, covering 20 diverse and culturally significant scenes from worldwide locations such as Cambridge Uni main buildings, the Pyramids, and the Forbidden City Palace. Compared to existing datasets, ours offers significantly larger scale and higher detail, uniquely suited for fine-grained 3D applications. Each scene contains an accurate spatial layout and comprehensive structural information, supporting detailed 3D reconstruction tasks. By reconstructing environments using these detailed images, our dataset supports multiple applications, including outputs in the widely adopted COLMAP format, establishing a novel benchmark for evaluating state-of-the-art large-scale Gaussian Splatting methods.The dataset's flexibility encourages innovations and supports model plug-ins, paving the way for future 3D breakthroughs. All datasets and code will be open-sourced for community use.

Figures (4)

• Figure 1: Overview of the 3D reconstruction pipeline and resulting model. (a) Original images of The Old Schools building. (b) Original images of Graduation Square. (c) Original images of the outside pavement area. (d) Reconstructed 3D model integrating multiple spatial views, highlighting key locations within the reconstructed environment.
• Figure 2: Main pipeline of our dataset. The workflow includes: (a) Raw image data processing (image acquisition, feature matching, sparse reconstruction, dense reconstruction, and mesh generation); (b) 3D Gaussian-Based Scene Rendering benchmarks; and (c) 3D model generation and evaluation (map alignment, camera pose estimation, and bundle adjustment).
• Figure 3: Overview of representative cultural heritage and urban environment datasets used in our benchmark.
• Figure 4: Qualitative comparison of reconstructed 3D scenes across multiple methods. The rows represent different reconstruction methods, while columns show three benchmark scenes.