USC BeoNav: A Large-scale, Long-term Offline Dataset for Visual Navigation

We collected a Large-scale visual navigation dataset over a span of 13 months across the entire USC campus. Unlike other popular dataset on Visual Navigation, our dataset contains sequences that are very diverse within a university setting. Due to the form factor of our robot, we were able to collect data from various settings like roads, outdoor lobby, ramps etc. that are usually present on a generic campus setting.

The entire dataset is collected over a course of 10 months and the visual data is recorded in a compressed format (.bag), that totals to about 114GB. We have data collected not only across diverse regions on campus, but also with different weather and lightening conditions. Some snippets of the dataset is shown in the video below.

The highlighted region on the following USC campus shows the area that was used to collect the data.

Each session in the dataset is recorded as a ROS Bag file. A bag file typically consists of a topic that contains a specific stream. In our dataset, a bag file consists of the following data streams and their respective sensor specifications:

 Note that IMU topics are missing in some bag files since we activated them after some data has been collected. Following are the .bag files that have been recorded along with the required details. You can find the license file here.

Sequence Name

If you use any of the above data for your work, please cite us using the following:


@online{beonav_2024,  author  = {Kiran Lekkala, Henghui Bao, Laurent Itti},  title   = {USCILab3D Dataset: A Large-scale, Long-term Outdoor 3D Dataset},  year    = {2024},  url     = {https://sites.google.com/usc.edu/beonav/home},  urldate = {2024-02-01},}