SJTU-ViSYS/M2DGR

What it solves

M2DGR provides a large-scale, multi-modal dataset specifically designed to benchmark and improve Simultaneous Localization and Mapping (SLAM) algorithms for ground robots. It addresses the lack of datasets that include rich sensory information and challenging real-world scenarios—such as complete darkness or elevators—where existing state-of-the-art SLAM solutions often fail.

How it works

The project provides 36 sequences (approximately 1TB of data) captured by a ground robot equipped with a comprehensive sensor suite. The data is synchronized and calibrated, and ground truth trajectories are provided using high-precision tools like motion capture devices, laser 3D trackers, and RTK receivers. The sensor suite includes:

• Vision: Six surround-view fish-eye cameras, one sky-pointing camera, a perspective color camera, an infrared camera, and an event camera.
• Lidar: A 32-beam Velodyne VLP-32C.
• Inertial/Positioning: Two IMUs and two GNSS receivers (including RTK).

Who it’s for

Researchers and developers working on ground robot navigation, sensor fusion, and SLAM algorithms who need a diverse and challenging benchmark to evaluate their systems' robustness and accuracy.

Highlights

• Rich Modality: Combines Lidar, RGB, infrared, event cameras, IMU, and GNSS data.
• Diverse Scenarios: Includes indoor and outdoor environments, including "corner cases" like lifts and total darkness.
• Comprehensive Benchmark: Used to evaluate and analyze the defects of various SOTA SLAM algorithms.
• Extensive Adoption: Serves as the foundation for numerous open-source SLAM projects and research papers.