Capturing, Reconstructing, and Simulating: the UrbanScene3D Dataset

This is the official repository of our ECCV22 dataset UrbanScene3D. Please refer to our project page for more information.

The whole dataset can be downloaded with Dropbox, Google Drive, or Baidu Yun.

Simulator

All the synthetic scenes and the reconstruction of real scenes are included in the simulator. Develop your Unreal Engine program to play with these scenes or install the AirSim plugin to capture your desired data. (Airsim uses PlayerStart as the origin.)

Open UrbanScene.uproject and rebuild the project. All the scenes are located in Contents/Maps in the Content Browser.

The models of the large-scale virtual scenes are not provided due to the copyright issue and the textured meshes can not be exported from Simulator(UE). We provide sampled points for them for evaluation purpose.

Path format

image_name,x,y,z,pitch,roll,yaw

The provided path files in our dataset are in UE (world) coordinate system, a left-handed, Z-Up coordinate system (in cm).

Each capture of synthetic scenes comes with a path file and a captured image set. The captured images of real scenes are recorded with GPS information.

Image capturing

For your convenience, we provide separate image-capturing programs (in this zip file) for the four synthetic scenes in the benchmark.

fov=60 resolution=6000*4000 for synthetic scenes. Real scenes come with different cameras, a camera calibration is needed to get the camera parameters accordingly.

Setup the path file and the directory of output files in ./cg_3_zuizhong/Saved/Config/WindowsNoEditor/Game.ini:

[targectory_path]
Path=path_to_path_file

[capture_path]
Path=direcotry_of_output_images

and run cg_3_zuizhong.exe to capture images.

Evaluation

Please download the compiled version for the evaluation on windows:x64.

evaluate_model.exe p1 path_to_recon p3 path_to_gt

p1,p3: "mesh" or "points"

The contained .pointcloud in real scene evaluation data is an encrypted format for the evaluation program. Download it and unzip it to data/. Delete triangles that don't belong to the buildings and pre-registrate the reconstructed model with the sparse point cloud PolyTech1k or ArtSci1k and then:

evaluate_model.exe p1 path_to_recon realscene data/artsci.pointcloud
#or
evaluate_model.exe p1 path_to_recon realscene data/polytech.pointcloud

UrbanScene3D-V1

The data of UrbanScene3D version 1 are available on Dropbox BaiduYun

or download separately from our NAS (more friendly for Chinese users, but maybe (quite) unstable sometimes):

• Residence
• Campus
• Sci-Art
• Square
• Hospital

Copyright

UrbanScene3D is publicly accessible for non-commercial uses only. Permission is granted to use the data only if you agree:

• The dataset is provided "AS IS". Despite our best efforts to assure accuracy, we disclaim all liability for any mistakes or omissions;
• All works that utilize this dataset including any partial use must include a reference to it. Please correctly cite our ECCV22 publication in research papers using the information provided below;
• You refrain from disseminating this dataset or any altered variations;
• You are not permitted to utilize this dataset or any derivative work for any commercial endeavors;
• We reserve all rights that are not explicitly granted to you.

BibTeX

@inproceedings{UrbanScene3D,
title={Capturing, Reconstructing, and Simulating: the UrbanScene3D Dataset},
author={Liqiang Lin and Yilin Liu and Yue Hu and Xingguang Yan and Ke Xie and Hui Huang},
booktitle={ECCV},
year={2022},
}