Synthetic 3D Lanes Dataset
Synthetic-3D-lanes is a dataset designed to train and validate methods for detection of lanes and center-lines in full 3D. It is a procedural randomly generated dataset of highway scenes with a large variability in road geometry and elevation. It is fully described in our ICCV`19 paper: 3D-LaneNet: End-to-End 3D Multiple Lane Detection . Here are some examples of images from the dataset:
Downloading the dataset. Download the dataset using this link. To download the dataset you'll have to have a Microsoft live account and request access for your account mail to the dataset by mailing dan.levi@gm.com
Dataset Details. see also README.txt in link
There are two distinct datasets:
Folder "paper_db" contains synthetic data used for ablation study described in the paper: 300K train images, 770 validation images, 5K test images.
Folder "with_ego_car_pos_variance" contains data with different ego car position and orientation (not necessarily centered and aligned with lanes): 117K Train images, 1K validation images.
Each image has a corresponding annotation file, with 3d landmark points (meters), 3d path points (meters), camera height (meters), camera pitch (degrees) and camera intrinsic parameters.
Coordinates are defined as: x-rightward, y-forward, z-upward, and camera intrinsic matrix defined accordingly.
See display_lanes_data.py for an example of reading and displaying data.
If you find this dataset useful in your research, please cite our paper:
@InProceedings{Garnett_2019_ICCV,
author = {Garnett, Noa and Cohen, Rafi and Pe'er, Tomer and Lahav, Roee and Levi, Dan},
title = {3D-LaneNet: End-to-End 3D Multiple Lane Detection},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}