Rapid growth of NAVER LABS is
owed to a network of researchers.
We are hoping to move forward together
by sharing our latest data.
This page introduces how we produce
high-precision data, and contains
basic information on the dataset.
Our technology is continuously advancing
at this very moment.
You can also check related stories here.
By using HD maps with sensor data, autonomous vehicles can perform accurate localization and plan routes effectively and safely. Therefore, HD maps are critical to the performance and safety of autonomous vehicles. This is why we are focusing on developing a new machine-readable HD map solution for on-road autonomous driving. NAVER LABS HD map is made using a Hybrid HD Mapping solution that organically integrates city-scale aerial images with MMS (mobile mapping system) data. While maintaining high precision, this can significantly reduce cost and time, compared to MMS vehicle generated maps. Our HD map, included in the open dataset, consists of the following three components.
NAVER LABS HD map provides 3D road layout
data extracted from aerial images. It contains road
structures essential for autonomous vehicles
including lanes, road signs, crosswalks, intersections
and speed bumps,
as well as their precise 3D
locations.
The HD map dataset contains 3D LiDAR point cloud
data of the surrounding road environment, scanned
with our MMS vehicle (R1).
Each point shows objects
on the road or the type of area. Dynamic objects,
such as cars or pedestrians, are automatically
detected and excluded from the point cloud.
The final component is visual features, extracted from key areas in the road environment. Visual feature data collected from deep learning models provide reliable matching and localization, as it is precise, discriminative and unchanging under different viewpoints.
During on-road autonomous driving, it is important
to identify the machine’s current location.
NAVER LABS’ localization uses internally developed
HD maps as a virtual sensor, and matches
information collected from diverse sensors such as
LiDAR, camera, inertia sensor and wheel encoder,
allowing precise and stable location estimation.
This works well even in gray areas with weak
GPS signals, such as urban centers and tunnels. We
provide raw data collected from sensors and
corresponding ground truth pose to evaluate
autonomous driving algorithm performance.
Visual localization is an important technology, providing the basis for indoor autonomous service robots and diverse location-based services. To facilitate related research, NAVER LABS is providing a dataset of various everyday spaces. This dataset contains different environmental components of actual spaces such as low ceilings, complex structures and crowds. NAVER LABS collected spatial data in spaces including department stores and subway stations full of people, and accumulated more than 130,000 images and point cloud data. Precise ground truth is provided by a data-processing pipeline that consists of LiDAR SLAM and SFM, and we simplified localization task evaluation by scanning each space repeatedly, at least twice.