MWAG: Multi-Season Wide-Area Air Ground Dataset for 3D Scene Reconstruction and Novel View Synthesis

Kshitij Singh Minhas, Qiao Wang, Niluthpol Chowdhury Mithun, Ben Southall, Supun Samarasekera, Rakesh Kumar

Student Lead Author Indication: No

Keywords: Computer Vision, 3D Reconstruction, Neural Rendering, Gaussian Splatting, Dataset

TL;DR: Multi-Season Wide-Area Air Ground Dataset for 3D Scene Reconstruction and Novel View Synthesis

Abstract: 3D scene reconstruction has seen significant advancements through methods like Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS), and is increasingly using generative AI models to improve the quality of reconstruction and novel view synthesis. However, existing datasets often lack the scale and diversity in complex, real-world outdoor environments. As a result, many of the state-of-the-art methods, including the generative models used in them, may not work well with real-world data. To further advance research in 3D reconstruction, neural rendering, and generative models on complex real-world data, we introduce the MWAG (Multi-season Wide-area Air Ground) dataset. MWAG features over 9,000 high-resolution RGB images captured from ground and aerial views across multiple seasons and differing environmental conditions. Each image includes highly accurate geodetic pose metadata, enabling tasks such as sparse-view 3D reconstruction, cross-view synthesis, and precise localization. The dataset supports multiple challenges in neural rendering, including handling environmental variations, training efficiency over expansive areas, and integrating multi-modal data for improved model completeness. We describe in detail our steps in data acquisition, processing, and alignment, to help the community create more diverse and challenging datasets to develop better methods and models in the future.

Submission Number: 39