No Wolf in the Meadow (NWM)

No Wolf in the Meadow (NWM) is a real-world outdoor image dataset of a flood control system in Vienna, Austria. The site covers approximately 100 × 200 meters and includes weir and flood control structures built from stone, wood, steel, and concrete, multiple river branches, and surrounding vegetation.

The dataset spans 8 recording days over 6.5 months (August 2025 – March 2026), capturing the site through seasonal vegetation changes, snow cover, flooding events, and active reinforcement and renovation work. All recordings were made with a consumer drone following pre-recorded flight plans, with 4K video exported as frames at regular intervals and downscaled to 1600 × 900 pixels. The dataset also includes detailed close-up recordings of specific scene areas. The dataset was created to evaluate ChronoFuseGS: Multi-Temporal Gaussian Fusion with Per-Splat Persistence and Change Visualization (currently under review)

The dataset is distributed as individual zip archives. Three types of archives are provided:

• nwm_<YYYY-MM-DD>.zip — Images and camera info (camera_info.json) for a single timestep.
• nwm_pre_trained_timestep_<YYYY-MM-DD>.zip — COLMAP model and pretrained 3DGS and PuP-3DGS models for a single timestep, used as input for ChronoFuseGS.
• nwm_<subset>.zip — Multi-temporal model trained and fused with ChronoFuseGS for a given subset (complete, autumn, flooding, snow, car).

We split the data into 11 timesteps in total. The two detail timesteps recorded at 2026-01-10 (2026-01-10_detail_1, 2026-01-10_detail_2) are only included in the Snow subset. The list of timesteps is shown below:

Each timestep's image set is split into training and test images. Every 8th image is held out as a test image; the remaining images are used for training. The type field in camera_info.json indicates the split for each image ("train" or "test").

We grouped these individual timesteps into the multi-temporal datasets of NWM, including the Complete dataset which covers all recording days, and smaller focused subsets. The table includes the number of registered images in the combined multi-temporal models.

nwm_<YYYY-MM-DD>/
└── data/
    ├── input/              # extracted drone video frames
    │   ├── <image>.png
    │   └── ...
    └── camera_info.json

nwm_pre_trained_timestep_<YYYY-MM-DD>/
├── data/
│   ├── images/             # images registered by COLMAP
│   ├── input/              # original extracted frames
│   ├── sparse/             # COLMAP sparse reconstruction
│   └── camera_info.json
├── output/
│   └── point_cloud/
│       ├── init_0/         # initial 3DGS model
│       └── iteration_30000/   # model after 30,000 refinement iterations
│           ├── point_cloud.ply
│           ├── activation.ply
│           └── camera.json
└── pup/                    # PuP-3DGS compressed models

nwm_<subset>/
├── data/
│   ├── images/
│   ├── input/
│   ├── sparse/
│   └── camera_info.json
└── output/
    └── point_cloud/
        ├── init_0/
        └── iteration_30000/
            ├── point_cloud.ply
            ├── activation.ply
            └── camera.json

We extracted GPS coordinates from the camera logs and provide per-image coordinate information in the corresponding camera_info.json file.

[
    {
        "image": "006_DJI_20250829153742_0593_D_1.png",
        "video_frame": 60,
        "src_video": "006_DJI_20250829153742_0593_D.MP4",
        "latitude": 48.206373,
        "longitude": 16.230399,
        "rel_alt": 20.0,
        "abs_alt": 317.406,
        "ecef": [4089114.54278001, 1190350.6104078596, 4732436.830631929],
        "registered": true,
        "type": "test",
        "t": 0
    },
    {
        "image": "006_DJI_20250829153742_0593_D_2.png",
        "video_frame": 180,
        "src_video": "006_DJI_20250829153742_0593_D.MP4",
        "latitude": 48.206436,
        "longitude": 16.2306,
        "rel_alt": 20.0,
        "abs_alt": 317.406,
        "ecef": [4089105.352015008, 1190363.495616108, 4732441.499502797],
        "registered": true,
        "type": "train",
        "t": 0
    },
    ...
]