README.md

---
pretty_name: WildCity
task_categories:
  - image-to-3d
  - depth-estimation
  - object-detection
tags:
  - autonomous-driving
  - 3d-reconstruction
  - gaussian-splatting
  - neural-rendering
  - lidar
  - rgb
  - city-scale
  - simulation
  - spatial-intelligence
  - multimodal
---

# WildCity Dataset

WildCity is a real-world city-scale multimodal dataset for street-view reconstruction, simulation, and spatial intelligence. It is collected from autonomous-driving fleet logs across multiple U.S. cities and contains surround-view RGB images, LiDAR, calibration, ego and sensor poses, object annotations, semantic masks, and processed reconstruction assets.

This repository hosts the **initial public release** of WildCity. This version does **not** include the full raw dataset. Instead, each released sequence provides the multimodal **raw data** and the **reconstruction-ready processed data** for the first ~22k keyframes of the corresponding log, with the processed COLMAP / 3DGS reconstruction offered at several keyframe scales.

For details about the dataset, benchmark, and reconstruction pipeline, please refer to our paper:

> **WildCity: A Real-World City-Scale Testbed for Rendering, Simulation, and Spatial Intelligence**  
> Xiangyu Han, Mengyu Yang, Jiaqi Li, Bowen Chang, Ziyu Chen, Hexu Zhao, Rahul Kumar Agrawal, Anthony Rodriguez, Rajani Acharya, Fiona Hua, Marco Pavone, Chen Feng, Yiming Li  
> Project page: `https://han-xiangyu.github.io/Wild-City/`  
> Paper: `https://arxiv.org/abs/2607.06838`

## Release Status

This is the **v0 initial release**.

Sequences are **anonymized**: each is identified by city and an index (e.g. `Ann-Arbor_02`), with capture dates and vehicle identifiers removed from all folder names.

Included in this release:

- The first ~22k keyframes for each released sequence.
- Raw (per-channel) surround-view RGB keyframe images.
- Raw LiDAR data.
- Per-sensor calibration (intrinsics + extrinsics).
- Ego poses and per-sensor poses (global and local frame).
- IMU data.
- Tracked-object annotations, object centers, and bounding boxes.
- Moving-object bounding-box overlays and depth visualizations.
- Preview surround-view videos.
- COLMAP / 3DGS-ready processed reconstructions at **multiple keyframe scales** (100, 500, 1k, 2k, 5k, 10k), each including undistorted images, semantic masks, LiDAR-camera depth, and a COLMAP sparse model.

Not included in this release:

- The complete raw logs.
- The full WildCity dataset.
- All city-scale benchmark segments.
- Pretrained reconstruction models or baseline checkpoints, unless explicitly added in a later release.

Future versions may expand the raw data coverage, processed segments, benchmark splits, and reconstruction outputs.

## Repository Structure

The repository is organized by **city**, then by **sequence**. Each sequence directory holds one `raw_data` archive plus several `processed_data_<N>_keyframes` archives (one per keyframe scale). All archives are `zstd`-compressed tarballs (`.tar.zst`).

```text
repo/
  Ann-Arbor/
    Ann-Arbor_02/
      raw_data.tar.zst
      processed_data_10k_keyframes.tar.zst
      processed_data_5k_keyframes.tar.zst
      processed_data_2k_keyframes.tar.zst
      processed_data_1k_keyframes.tar.zst
      processed_data_500_keyframes.tar.zst
      processed_data_100_keyframes.tar.zst
  Arlington/
    Arlington_01/
      ...
  Atlanta/
    Atlanta_02/
      ...
  Eden-Prairie/
    Eden-Prairie_01/
      ...
  Grand-Rapids/
    Grand-Rapids_01/
      ...
  Miami/
    Miami_02/
      ...
```

Six cities are included: **Ann-Arbor, Arlington, Atlanta, Eden-Prairie, Grand-Rapids, Miami**. Each city currently provides one sequence; the `<City>_NN` index leaves room for additional sequences per city in future releases.

Each sequence has two archive families:

- **`raw_data.tar.zst`** — the multimodal raw data and metadata for the extracted keyframes.
- **`processed_data_<N>_keyframes.tar.zst`** — a self-contained COLMAP / 3DGS-ready reconstruction built from the first **N** camera keyframes (a keyframe is one synchronized 6-camera capture). The scales are nested prefixes: `100 ⊂ 500 ⊂ 1k ⊂ 2k ⊂ 5k ⊂ 10k`, so all start from the same first frame. Use the smallest scale that fits your compute for quick experiments, or `10k` for the full sequence.

## Archive Contents

### `raw_data.tar.zst`

Extracts to a `raw_data/` folder:

```text
raw_data/
  images/                            # surround-view keyframe images, per camera channel
    image_lucid_fc_jpeg/ fl/ fr/ bc/ lc/ rc/
  lidar/                             # per-keyframe LiDAR point clouds
  calibration/                       # camera intrinsics + sensor extrinsics
  depth_visualization/               # depth preview images (inspection only)
  object_overlays/                   # moving-object bounding-box overlays on images
  videos/                            # preview surround-view videos
  imu.csv                            # IMU measurements
  keyframes_pose.csv                 # ego poses
  keyframes_sensor_poses.csv         # global per-sensor poses
  keyframes_sensor_poses_local.csv   # local-frame per-sensor poses
  keyframes_tracked_objects.csv      # tracked-object annotations
  keyframes_object_bboxes.jsonl      # per-keyframe object bounding boxes
  keyframes_object_centers.jsonl     # per-keyframe object centers
  undistortion_intrinsics.json       # intrinsics used for image undistortion
```

- `images/` — the original surround-view camera images for the extracted keyframes, organized in one subfolder per camera channel. These are the source images prior to the 3DGS-specific undistortion applied in the processed data. The six channels are front-center (`fc`), front-left (`fl`), front-right (`fr`), back-center (`bc`), left-center (`lc`), and right-center (`rc`).
- `lidar/` — raw LiDAR point clouds for the released keyframes.
- `calibration/` — per-sensor calibration defining the geometric relationship between cameras, LiDAR, and the ego-vehicle frame.
- `depth_visualization/` — depth preview images for quick inspection/debugging, not for metric evaluation.
- `object_overlays/` — visualizations of moving-object bounding boxes overlaid on images, useful for inspecting dynamic-object annotations and filtering behavior.
- `videos/` — preview surround-view videos of the sequence.
- `imu.csv` — raw IMU measurements.
- `keyframes_pose.csv` — ego poses for the keyframes.
- `keyframes_sensor_poses.csv` / `keyframes_sensor_poses_local.csv` — per-sensor poses in global and local frames.
- `keyframes_tracked_objects.csv` — tracked-object annotations.
- `keyframes_object_bboxes.jsonl` / `keyframes_object_centers.jsonl` — per-keyframe object bounding boxes and centers.
- `undistortion_intrinsics.json` — camera intrinsics used for the undistortion in the processed reconstruction.

### `processed_data_<N>_keyframes.tar.zst`

Extracts to a `processed_data_<N>_keyframes/` folder — a COLMAP / 3DGS-ready reconstruction using the first `N` keyframes. Intended to be directly usable by COLMAP-style and 3D Gaussian Splatting pipelines.

```text
processed_data_<N>_keyframes/
  images/            # undistorted images, flat: <channel>_<index>.jpg, all 6 cameras
  masks/
    dynamic/         # moving / potentially-dynamic object masks
    ground/          # road / ground masks
    sky/             # sky masks
  depth_map_gt/      # LiDAR-camera reprojection depth, one .npz per image
  sparse/
    0/               # COLMAP binary model: cameras.bin, images.bin, points3D.bin
```

- `images/` — undistorted images prepared for reconstruction (the preferred input for 3DGS training), named `<channel>_<index>.jpg`.
- `masks/` — semantic masks for region-aware reconstruction: `dynamic/` (moving-object filtering), `ground/` (ground regularization), and `sky/` (sky modeling). Masks are automatically generated.
- `depth_map_gt/` — LiDAR-camera reprojection depth maps providing sparse depth supervision derived from LiDAR projected into camera views.
- `sparse/0/` — COLMAP sparse model in the standard binary layout (`cameras.bin`, `images.bin`, `points3D.bin`); the point cloud is initialized from accumulated LiDAR.

## Extracting an Archive

Archives are `zstd`-compressed tarballs. With GNU tar (which supports `--zstd`):

```bash
tar --zstd -xf Ann-Arbor/Ann-Arbor_02/processed_data_5k_keyframes.tar.zst
```

Or decompress explicitly with `zstd`:

```bash
zstd -dc Ann-Arbor/Ann-Arbor_02/raw_data.tar.zst | tar -xf -
```

Each archive extracts a single top-level folder (`raw_data/` or `processed_data_<N>_keyframes/`).

## Recommended Usage

### Raw-data inspection

Extract `raw_data.tar.zst` and use:

```text
raw_data/images/
raw_data/lidar/
raw_data/calibration/
raw_data/keyframes_pose.csv
raw_data/keyframes_sensor_poses.csv
raw_data/keyframes_tracked_objects.csv
```

These are suitable for inspecting raw multimodal sensor data, poses, calibration, and object annotations.

### 3DGS reconstruction

Extract a `processed_data_<N>_keyframes.tar.zst` and use:

```text
processed_data_<N>_keyframes/images/
processed_data_<N>_keyframes/masks/
processed_data_<N>_keyframes/depth_map_gt/
processed_data_<N>_keyframes/sparse/0/
```

Start with a small scale (e.g. `processed_data_1k_keyframes`) for quick iteration, and use `processed_data_10k_keyframes` for the full sequence.

## Dataset Versioning

### v0

Initial release.

- Sequences are anonymized as `<City>_NN` (dates and vehicle ids removed).
- `raw_data` covers the first ~22k keyframes; processed reconstructions are provided at 100 / 500 / 1k / 2k / 5k / 10k keyframes.
- Full raw logs are not included.
- The release focuses on providing a usable subset with both raw multimodal data and 3DGS-ready processed data.

## Notes and Limitations

This initial release is intended to support reproducible research on city-scale reconstruction and neural rendering. It should not be interpreted as the complete WildCity dataset.

Known limitations of this version:

- Only the first ~22k keyframes are included for each released sequence, and the processed reconstructions cover at most 10k keyframes.
- Processed keyframe scales are nested prefixes of one another (they share the same starting frame), not independent samples across the full route.
- Some full raw logs and full benchmark segments are not yet released.
- Semantic masks are automatically generated and may contain boundary errors or rare category mistakes.
- Poses and calibration are provided for research use, but residual pose noise may still exist in long-horizon real-world logs.

## Citation

If you use this dataset, please cite:

```bibtex
@inproceedings{han2026wildcity,
  title     = {WildCity: A Real-World City-Scale Testbed for Rendering, Simulation, and Spatial Intelligence},
  author    = {Han, Xiangyu and Yang, Mengyu and Li, Jiaqi and Chang, Bowen and Chen, Ziyu and Zhao, Hexu and Agrawal, Rahul Kumar and Rodriguez, Anthony and Acharya, Rajani and Hua, Fiona and Pavone, Marco and Feng, Chen and Li, Yiming},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026}
}
```

Please update the BibTeX entry according to the official proceedings version.

## License

This dataset is released under the **Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0)**, unless otherwise specified.

Under this license, you may use, share, and adapt the dataset for **non-commercial research and educational purposes**, provided that you:

- give appropriate credit to the WildCity authors;
- indicate whether changes were made;
- do not use the dataset for commercial purposes; and
- distribute any adapted material under the same or a compatible license.

See the `LICENSE` file in this repository for the license notice and the official Creative Commons legal code.

## Contact

For questions about the dataset, please contact:

```text
xiangyuhan615@gmail.com
```