USCILab3D Dataset: A Large-scale, Long-term Outdoor Dataset
Introduction
We are dedicated to building a unique dataset that goes beyond traditional mapping methods. By harnessing the power of point clouds and robot poses, we are create a comprehensive outdoor dataset that consists of large, long-term semantically annotated pointclouds. It is a dataset included by Open X-Embodiment and has been used in navigation tasks.
Multi-view, Multi-modal, Large-scale
With an extensive repository comprising over 10 million images and 600,000 point cloud files, our dataset serves as a cornerstone for cutting-edge research and innovation.
At the intersection of computer vision, robotics, machine learning, and beyond, our platform empowers researchers to embark on ambitious projects and push the boundaries of knowledge. Whether you're delving into object recognition, scene understanding, or 3D reconstruction, our dataset provides the raw materials necessary for groundbreaking discoveries.
Semantic Annotations using 2D-3D Correspondences
We use GPT4 to create the labels in the images and use as input for Grounded SAM, then we align the 2D and 3D points by projection to create the 3D semantic labels.
Dataset collection
We conducted an extensive data collection endeavor spanning 13 months throughout the entirety of the University of Southern California (USC) campus. Our dataset encompasses diverse sequences characteristic of a university environment. Leveraging the form factor of our robotic platform, we systematically gathered data from a multitude of settings, including but not limited to roads, outdoor lobbies, and ramps typically found within a generic campus landscape. We We also release the robot dataset and is part of Open X-Embodiment that could be used for navigation tasks.
Dataset information
Above: We specify the total number of images and pointclouds that are collected during a session on a specific day, along with number of sessions on a given day.