gpdaniels/slam-datasets

SLAM Datasets

A collection of commonly used Simultaneous Localization and Mapping (SLAM) datasets being used in the development of ZeroSLAM.

All datasets have been converted to (a subset of) the MCAP format and into a standard coordinate system for calibration and poses.

Overview

This repository brings together several established robotics and computer-vision datasets in a consistent MCAP-based format for use with robotics, SLAM, visual odometry, localization, mapping, and sensor-fusion pipelines.

The collection currently contains:

• ETH3D SLAM
• EuRoC MAV
• KITTI Odometry
• TUM RGB-D

Scene format

A scene is one mcap file, [dataset]/[scene].mcap, holding something like:

• Raw image/depth messages in lz4 compressed chunks.
• Camera intrinsics as one camera info message per frame.
• IMU messages.
• A ground truth trajectory on /tf as root -> ego.

Sensors are topics named /sensor/[type]_[01-99] (image_01, image_02, ... for cameras; imu_01, ... for imus; lidar_01, gnss_01...) and each sensor's frame carries its full name (e.g. sensor/image_01).

Transforms are stored in /tf following root -> ego -> sensor/[name] with the root -> ego transform being the ground truth. A per topic message ego -> sensor/[name] transform poses each sensor on it with the calibration extrinsics (so extrinsics could change over time).

Scene mcap files can be viewed directly in a browser with web-viewers e.g. Lichtblick.