360Loc: A Dataset and Benchmark for Omnidirectional Visual Localization with Cross-device Queries
Huajian Huang*1, Changkun Liu*1, Yipeng Zhu1, Hui Cheng2,
Tristan Braud1 and Sai-Kit Yeung1
* equal contribution
The Hong Kong University of Science and Technology1, Sun Yat-Sen University2
In Proceedings of Computer Vision and Pattern Recognition Conference (CVPR), 2024
To save the store space and facilitate data access, we only store the original 360-degree images of query and inference. You can use this tool to generate required images in fisheye and pinhole cameras.
git clone https://github.com/HuajianUP/360Loc.git
pip install json tqdm yaml numpy opencv-python requests
You can download the 360Loc dataset link and unzip them in the folder ./360Loc
.
360Loc
├── atrium
│ ├── camera_config
│ ├── pose
│ ├── mapping
│ ├── query_360
│ ├── query_pinhole
│ ├── query_fisheye1
│ ├── query_fisheye2
│ └── query_fisheye3
├── ....
│
└── piatrium
├── camera_config
├── ....
└── query_fisheye3
And then run the below command to process the data.
python process.py --dir PATH_TO_360LOC_DATASET
We also provide a script to download and process the 360Loc dataset automatically.
python onekey.py --dir PATH_TO_SAVE_360LOC_DATASET
We provide both json poses files and txt poses files.
For .json poses files:
We give 4*4 matrix
image_name: [R|T]
, T
are camera to world coordinates.
For .txt poses files:
image_name x y z qw qx qy qz
.
x, y, z
are camera to world coordinates.
Our 6DoF poses follow the COLMAP coordinate system.
@inproceedings{360Loc,
title = {360Loc: A Dataset and Benchmark for Omnidirectional Visual Localization with Cross-device Queries},
author = {Huang, Huajian and Liu, Changkun and Zhu, Yipeng and Cheng Hui and Braud, Tristan and Yeung, Sai-Kit},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year = {2024}
}