Xiaohang Yu · Zhengxian Yang · Shi Pan · Yuqi Han · Haoxiang Wang · Jun Zhang · Shi Yan · Borong Lin · Lei Yang · Lu Fang · Tao Yu
Under Submission
| Scene | Camera | pos | img | volume | 5K JPEGs |
|---|---|---|---|---|---|
| RuziNiu Statue | 41 | 38 | 1,558 | 11.52 m³ | 19 GB |
| LizhaoJi Building | 41 | 21 | 861 | 6.83 m³ | 9 GB |
| Architecture | 40 | 32 | 1,295 | 12.57 m³ | 13 GB |
| Center Garden | 40 | 53 | 2,109 | 23.95 m³ | 23 GB |
| Square1 | 40 | 40 | 1,588 | 8.37 m³ | 15 GB |
| Square2 | 40 | 47 | 1,876 | 11.59 m³ | 18 GB |
| Flagstaff | 40 | 38 | 1,518 | 8.04 m³ | 15 GB |
| Office1 | 40 | 43 | 1,710 | 13.14 m³ | 15 GB |
| Office2 | 40 | 89 | 3,550 | 39.72 m³ | 30 GB |
| Total | 157 GB |
* volumecolumn indicates the space size with at least 10 camera within a range of 1 m³
* Above are all valid data which we have undistorted and tested on Metashape. If you want to acquire raw data, please contact us.
| RuziNiu Statue | LizhaoJi Building | Architecture |
|---|---|---|
 |
 |
 |
| Center Garden | Square1 | Square2 |
|---|---|---|
 |
 |
 |
| Flagstaff | Office1 | Office2 |
|---|---|---|
 |
 |
 |
All images in the dataset were taken with our self-designed multi-camera device named "Compound Eye" (as specified in the overview table). Capture rig consists of 40-41 GoPro HERO10 cameras mounted on a remote-controlled car.
The Den-SOFT dataset is hosted on one Drive, and can be explored with any browser.
Our scene capture data is organized following this structure:
1.41 cameras: RuziNiu Statue and LizhaoJi Building
│
├── images # Scene images with 5K resolution
│ ├── 1(1) # First camera (first pose)
│ ├── 1(2)
│ ├── 1(...) # More images
│ ├── 2(1) # Second camera (first pose)
│ ├── [...] # More images
│ ├── 41(38) # Last camera (last pose)
2.40 cameras: All other scene datas
│
├── images # Scene images with 5K resolution
│ ├── Square1-1 # Scene name - first pose
│ ├── [...] # More images
3.Metashape Reference: camera poses and point cloud
│
├── Architecture # Scene name
│ ├── camera.xml # Camera poses
│ ├── point_cloud.ply # Point cloud correspond to the scene
- We provide images at 5K resolution(Width:5568, Height:4176). If you want to use Metashape to get sparse point cloud reconstruction or camera poses, you can use the
camera.xml,point_cloud.plyas references. Some of the images were removed due to involving irrelevant elements in the scene or blurry images. It would be greatly appreciated if you have a better method that can use all images to estimate camera pose and reconstruct point clouds.
The following videos show the training result of our datasets using 3DGS.
3DGS_Demo10M.mp4
The following video shows the comparison between training results and Ground truth.
side_by_side_comparisonV_10M.mp4
In order to give you a more intuitive view of the sampling density when we collect scene data, we have visualized the viewpoint density within a unit sphere using Metashape and mayavi. The sampling density from blue to red indicates sparse to dense.
Square1_density.mp4
Office1_density.mp4
If you use any data from this dataset or any code released in this repository, please cite the Den-SOFT paper.
@InProceedings{Den-SOFT,
author = {Xiaohang Yu and
Zhengxian Yang and
Shi Pan and
Yuqi Han and
Haoxiang Wang and
Jun Zhang and
Shi Yan and
Lei Yang and
Lu Fang and
Tao Yu },
title = {{Den-SOFT}: Dense Space-Oriented Light Field Dataset for 6-DOF Immersive Experience},
year = {2024},
url = {https://metaverse-ai-lab-thu.github.io/Den-SOFT/},
}Creative Commons Attribution-NonCommercial (CC BY-NC) 4.0, as found in the LICENSE file.