MV2 Dataset

MV2 is a multi-view and multi-vehicle urban driving dataset designed for research in novel view synthesis, neural rendering, 3D reconstruction, cross-view scene understanding, and autonomous-driving perception. The dataset contains synchronized image sequences captured from multiple viewpoints, including ground vehicles and aerial views, along with camera parameters required for geometry-aware learning and rendering.

The dataset is released for academic and research use. The full dataset is being made available here, and additional annotations such as segmentation masks for dynamic objects may be released in upcoming updates.

Dataset Details

Dataset Description

MV2 focuses on challenging real-world urban driving scenes with large viewpoint changes across vehicles and platforms. It is intended to support research on:

• Cross-view novel view synthesis
• Multi-vehicle and multi-camera scene reconstruction
• Ground-to-aerial and aerial-to-ground rendering
• Geometry-aware neural rendering
• Autonomous-driving scene understanding
• Dynamic-scene analysis in urban traffic environments

Unlike single-ego-vehicle driving datasets, MV2 includes multiple observation platforms, making it useful for studying extreme-baseline view synthesis and shared world modeling.

Dataset Sources

• Repository: TODO: add Hugging Face dataset URL
• Project page: TODO: add MV2 project page URL
• Paper: TODO: add paper/arXiv link
• Code: TODO: add GitHub repository link

Dataset Structure

The dataset is organized at the sequence level. Each sequence contains multiple views/platforms such as car, drone, scooty, and additional car-view variants.

A typical structure is: