MegaScenes: 430k Scenes with 3D Reconstructions and 2 Million Images

The MegaScenes Dataset is a collection of around 430,000 scenes, featuring over 100,000 structure-from-motion reconstructions and over 2 million registered images. It includes a diverse array of scenes such as minarets, building interiors, statues, bridges, towers, religious buildings, and natural landscapes. The dataset was contributed by Cornell University and is available under a CC-BY-4.0 license.

Use Cases

• Training and benchmarking 3D reconstruction models based on the over 100,000 structure-from-motion reconstructions.
• Developing algorithms for visual localization and mapping based on the 2 million registered images.
• Researching view synthesis and novel view generation based on images captured under varying conditions.
• Analyzing scene representation and feature matching across diverse scenes like buildings, statues, and landscapes.

Strengths

• Large scale with approximately 430,000 scenes.
• Includes over 100,000 3D reconstructions and over 2 million images.
• Contains diverse scene types, including man-made structures and natural landscapes.
• Images were captured under varying conditions of time, weather, illumination, and camera intrinsics.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for specific data components is unknown, which may limit suitability assessment.
• Last update date is unknown; freshness unverified.

Provenance

Source

Cornell University (https://www.cs.cornell.edu/~snavely/)

Collection Method

Collection of structure-from-motion reconstructions and registered images from various sources and conditions.

Time Range

null

Freshness

Last updated: unknown

Geography

null