Zqlin/TwinWorld_Datasets

TwinWorld 2026 Dataset

Drone-captured urban building scenes for the TwinWorld 2026 (ECCV) challenge: Gaussian Splatting-based novel-view synthesis and 3D reconstruction (geometry and semantics). The challenge data is preprocessed from two real-world collections:

• TUM, from TUM2TWIN
• Gold Coast, from GT-LOD3-Benchmark

TUM2TWIN subsets used in this challenge. Red = development scenes, blue = final-testing scenes.	Local visualization of the Gold Coast data.

What's included

Twinworld_Datasets/
├── Data_TUM/
│   └── scene_000 .. scene_008/
│       ├── train/
│       │   ├── images/*.JPG
│       │   └── sparse/0/
│       │       ├── cameras.txt
│       │       ├── images.txt
│       │       ├── points3D.txt
│       │       └── points3D.ply        # initialization point cloud
│       ├── test/
│       │   ├── images/*.JPG            # scene_000-003 only
│       │   └── sparse/0/
│       │       ├── cameras.txt
│       │       └── images.txt          # poses only, always provided
│       └── 3d_gt/
│           └── point_cloud.ply         # scene_000-003 only, x, y, z
├── Data_Goldcoast/
│   └── scene_009 .. scene_012/         # same layout as above
│       └── 3d_gt/point_cloud.ply       # scene_009-010 only, x, y, z, classification
└── script/
    └── render_test_poses.py

train/ and test/ inside one scene share the same world coordinate system (one joint COLMAP reconstruction per scene), so a model trained on train/ can be rendered directly at the poses listed in test/sparse/0/. Camera intrinsics/extrinsics use the standard COLMAP text format (cameras.txt, images.txt); points3D is triangulated from all images (train and test alike) and is only provided under train/sparse/0/.

The ground-truth point cloud (where provided) carries x, y, z for TUM scenes, plus an integer classification for Gold Coast scenes:

ID	Class
0	ground
1	wall
2	roof
3	window
4	other
255	ignore

Development scenes vs. final testing scenes

	TUM	Gold Coast	What you get
Development	scene_000–scene_003	scene_009–scene_010	Everything: train/, test/images + test/sparse/0, and 3d_gt/point_cloud.ply.
Final testing	scene_004–scene_008	scene_011–scene_012	train/ and test/sparse/0 (camera poses only). No test/images and no 3d_gt/point_cloud.ply are included.

For final-testing scenes, test/sparse/0/images.txt still lists the full camera pose for every held-out frame, so you always know exactly which viewpoints to render, even though you never receive the photos or geometry those poses were held out from. Development scenes give you everything locally so you can self-check before relying on the final-testing scenes.

Training example

Using graphdeco-inria/gaussian-splatting (any other method/codebase works the same way, as long as it consumes COLMAP-format input):

git clone --recursive https://github.com/graphdeco-inria/gaussian-splatting
cd gaussian-splatting

python train.py -s <path-to>/Twinworld_Datasets/Data_Goldcoast/scene_009/train -m <output_dir>

Do not pass --eval: the train/test split is already fixed by this dataset's own train//test/ folders, so train.py should not carve out its own held-out views from train/.

Rendering test views

script/render_test_poses.py renders a trained model at this dataset's test camera poses. It targets the vanilla gaussian-splatting repo above, so copy it into that repo's root (next to train.py) and run it from there:

python render_test_poses.py \
    --model_ply <output_dir>/point_cloud/iteration_30000/point_cloud.ply \
    --camera_pose_dir <path-to>/Twinworld_Datasets/Data_Goldcoast/scene_009/test/sparse/0 \
    --output_dir <render_dir>/scene_009

This writes <render_dir>/scene_009/rgb/<frame_id>.png, one image per pose in test/sparse/0/images.txt, named after that pose's own image name. If you used a different method/codebase, adapt the model-loading and rendering calls to your own repo, keeping the same camera-pose parsing (cameras.txt/images.txt) and output naming.

Checking your results with metrics.py

For development scenes, where the real test photos are provided locally, you can self-check PSNR/SSIM/LPIPS with gaussian-splatting's own metrics.py. It expects renders and reference photos arranged in its own renders//gt/ layout rather than the flat rgb/ folder render_test_poses.py produces, so a quick reorganization is needed first. See metrics.py's own usage notes for the exact layout it expects. Final-testing scenes have no local ground truth, so scores for those only come back after official evaluation.

References

Wysocki, Olaf, et al. "TUM2TWIN: Introducing the large-scale multimodal urban digital twin benchmark dataset." ISPRS Journal of Photogrammetry and Remote Sensing 232 (2026): 810–830.

Kim, Han Sae, et al. "GT-LOD3: LOD3 Semantic 3D Building Reconstruction Benchmark Dataset." ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences 11 (2026): 293–302.