# Download and assemble WildCity scenes These instructions describe the full-log reconstruction `packages/` layout. Camera videos and their metadata will be published separately at a later date. The initial release's `raw_data.tar.zst` and `processed_data_*_keyframes.tar.zst` layout uses different instructions. Do not mix files from the two layouts in one extraction directory. ## 1. Download only the required sequence The Hugging Face repository is `Neptune615/Wild-City`. Once this release is published, select one sequence with `snapshot_download`: ```bash python -m pip install huggingface_hub ``` ```python from huggingface_hub import snapshot_download snapshot_download( repo_id="Neptune615/Wild-City", repo_type="dataset", local_dir="wildcity_download", allow_patterns=[ "README.md", "LICENSE", "docs/*", "inventory/release.json", "packages/Ann-Arbor/Ann-Arbor_01/*", "code/processing_code*", "code/assembly_interface.json", ], ) ``` For a reproducible experiment, also pass `revision=""` after choosing the published revision. The `allow_patterns` example downloads **all assets for one log**, which can still be hundreds of GB. Individual shards can be selected by filename after consulting their `.files.jsonl.gz` inventories; each row lists a relative member path, byte size, and SHA-256. A frame ID need not reside in an identically numbered shard in different cameras or modalities. API reference: [Hugging Face download guide](https://huggingface.co/docs/huggingface_hub/guides/download). ## 2. Verify downloaded archives The following verifies all shards and compressed member inventories of the selected log. It reads every downloaded archive; allow time proportional to the downloaded size. ```python import hashlib import json from pathlib import Path package = Path("wildcity_download/packages/Ann-Arbor/Ann-Arbor_01") manifest = json.loads((package / "manifest.json").read_text()) for shard in manifest["shards"]: for name, expected in [ (shard["name"], shard["sha256"]), (shard["manifest"], shard["manifest_sha256"]), ]: path = package / name digest = hashlib.sha256() with path.open("rb") as stream: for block in iter(lambda: stream.read(4 * 1024**2), b""): digest.update(block) if digest.hexdigest() != expected: raise ValueError(f"Checksum mismatch: {path}") print("All downloaded shards and inventories match.") ``` Code archive checksums are in `code/processing_code.receipt.json`. ## 3. Extract one log Install GNU tar and the `zstd` executable, then run in Bash: ```bash set -euo pipefail package=wildcity_download/packages/Ann-Arbor/Ann-Arbor_01 data=extracted/Ann-Arbor_01 mkdir -p "$data" for archive in "$package"/*.tar; do tar -xf "$archive" -C "$data" done for archive in "$package"/*.tar.zst; do tar --zstd -xf "$archive" -C "$data" done ``` Use a fresh extraction directory per log. Shards contain paths relative to that log's extraction root, **not** another top-level `raw_data/` directory: ```text extracted/Ann-Arbor_01/ images/image_lucid_fc_jpeg/image_lucid_fc_jpeg_00000.jpg images/image_lucid_fl_jpeg/... ... lidar/keyframe_00000_velodyne_fc_points_color.npz metadata/ calibration/mcm_to_colmap_calibration.json undistortion_intrinsics.json keyframes_pose.csv keyframes_sensor_poses.csv ... masks/dynamic/.png masks/ground/.png masks/sky/.png selections/current.json selections/coverage.json ``` Point-cloud filenames may also use `_points.npz`; the assembly tool supports both forms. Keep the supplied names. `current.json` is a JSON list of original keyframe IDs for which masks are supplied; `coverage.json` records coverage and known missing data. ## 4. Install the small assembly environment For assembly from already decoded assets, a CPU environment with NumPy and Pillow is sufficient. The following versions were tested in an isolated Python 3.10 environment: ```bash python3 -m venv wildcity_assembly_env source wildcity_assembly_env/bin/activate python -m pip install numpy==2.2.6 pillow==12.0.0 mkdir -p processing_code tar -xzf wildcity_download/code/processing_code.tar.gz -C processing_code python processing_code/assemble_colmap.py --help ``` The bundle also contains the original processing pipeline and third-party sources. Its broad `requirements.txt` and machine-specific runner examples describe upstream log decoding and GPU processing, not this minimal assembly workflow. Re-running original MCM decoding requires the May toolkit and original recordings; generating new segmentation/anonymization outputs requires the corresponding dependencies and model weights, which are not distributed here. ## 5. Assemble a masked subset Select the first 100 IDs from the existing spatial selection: ```python import json from pathlib import Path frames = json.loads(Path("extracted/Ann-Arbor_01/selections/current.json").read_text()) Path("frames.json").write_text(json.dumps(frames[:100])) ``` Then build a new scene: ```bash python processing_code/assemble_colmap.py \ --data-root extracted/Ann-Arbor_01 \ --output-root scenes/Ann-Arbor_01_first100 \ --frames-json frames.json \ --points-decimation 4 \ --ground-downsample-factor 6 ``` The output directory must **not exist** and must be **outside** the extracted input directory. All six cameras are used by default. For fewer cameras, add, for example, `--channels IMAGE_LUCID_FC_JPEG IMAGE_LUCID_FL_JPEG`. The scene contains copied images, masks, `sparse/0/{cameras,images,points3D}.{bin,txt}`, and `assembly.json`. Intrinsics and poses are supplied by the dataset; sparse geometry comes from LiDAR. No image matching or bundle adjustment is run. ## 6. Choose other full-log frames For any valid original frame IDs, explicitly omit masks if those frames are not in the existing selection: ```bash python processing_code/assemble_colmap.py \ --data-root extracted/Ann-Arbor_01 \ --output-root scenes/Ann-Arbor_01_custom \ --frames 0 10 20 \ --without-masks \ --points-decimation 8 ``` Choose IDs present in `metadata/keyframes_pose.csv`. A JSON list can contain any available noncontiguous subset from that same log. Spatially select IDs using the released poses, then pass the resulting list through `--frames-json`. `--points-decimation` and `--ground-downsample-factor` control geometry density, not which camera frames are selected. `--max-points` additionally limits exported points. Combining separate logs requires an external alignment procedure. The assembly entrypoint does not infer cross-log transforms. ## 7. Miami_01 IMU `metadata/imu.csv` contains the original KVH messages for Miami_01. The `utime` column is the payload timestamp in microseconds; `rotation_x/y/z` are **delta angles in radians** (`rotation_type=1`), and `accel_x/y/z` are in m/s². Values remain in the native sensor axes, with no resampling or coordinate transform. Do not interpret delta angles as angular velocity. `metadata/imu_info.json` records sample counts, timestamp checks, units, and the mounting geometry from `param-matrix.config`. Reprocessing this log with Stage 1 requires `--imu-channel KVH`; its default channel `IMU` does not match this recording.