metadata/raw_dataset_schema.md

# Scene2Wave dataset schema v1

Scene2Wave `1.0.1` contains 100 one-second multimodal CARLA/Sionna samples. The
machine-readable JSON Schema is `metadata/json_schema.json`; this document
defines file roles, units, synchronization rules, and minimal reading examples.

## Sample layout

```text
data/main_training/<Town>/<state>/<profile>/<scenario>/
├── sample_metadata.json
├── carla/
│   ├── alignment_index.json
│   ├── cav_1/
│   ├── rsu_1/
│   ├── birdview/
│   └── scenes/
└── sionna/
```

Generator cache keys, internal asset manifests, visualizations, and derived
videos are not release data and are deliberately excluded.

## JSON and JSONL documents

### `dataset_info.json`

Release-level identity, provenance, distribution counts, byte/file counts, and
the explicit inclusion/exclusion policy. Validate it with JSON Schema definition
`datasetInfo`.

### `metadata/samples.jsonl`

The authoritative sample index. Each non-empty line is one independent JSON
object matching `sampleIndexRecord`.

| Field | Type | Unit / meaning |
|---|---|---|
| `sample_id` | string | Unique sample identifier. |
| `relative_path` | string | Dataset-root-relative sample directory. |
| `town` | string | CARLA map. |
| `state` | string | `static`, `10kmh`, `20kmh`, `40kmh`, or `60kmh`. |
| `profile` | string | Dataset/channel profile bucket. |
| `scenario` | string | Generator scenario name. |
| `valid_cir` | boolean | Whether the formal CIR sample passed validity checks. |
| `contains_empty_cir` | boolean | Whether any selected CIR frame is empty. |
| `carrier_frequency_hz` | number | RF carrier frequency in Hz. |
| `bandwidth_hz` | number | Occupied bandwidth in Hz. |
| `subcarrier_spacing_hz` | number | OFDM subcarrier spacing in Hz. |
| `num_subcarriers` | integer | Number of CSI subcarriers. |
| `csi_sampling_rate_hz` | number | CIR/CSI temporal sampling rate in Hz. |

### `sample_metadata.json`

A compact user-facing record matching `sampleMetadata`. It intentionally keeps
only stable fields needed for filtering, interpretation, reproduction, and the
browser demo.

- `dataset`: subset/profile/role/split/Town/state labels.
- `radio`: carrier, bandwidth, subcarrier spacing/count, and temporal rate.
- `channel`: validity, outage, LOS/NLOS, path-count, delay-spread, and material
  profile summaries.
- `geometry`: RSU/CAV positions and mean link distance in CARLA world metres.
- `ray_tracing`: the public Sionna RT computation summary.
- `sensor_context`: birdview calibration and route centre needed to place
  overlays. It is not a generator cache.
- `provenance`: stable hashes only; no workstation path.
- `paths`: sample-relative locations of CARLA, Sionna, and alignment data.

Coordinates under `geometry` use the CARLA world frame and metres. The file does
not replace the per-frame YAML pose records under `carla/cav_1/` and
`carla/rsu_1/`.

### `carla/alignment_index.json`

The authoritative mapping between the 2 kHz geometry/CIR/CSI clock and lower
rate sensors. Validate it with `alignmentIndex`.

- `geometry.frames[]` maps `geometry_index` to CARLA `frame_id` and `time_s`.
- `streams[]` identifies every sensor stream and its actual observed frames.
- `observed_frames[].nearest_geometry_frame_id` is the geometry frame associated
  with the sensor callback.
- `planned_samples[]` records the requested schedule and is retained for
  acquisition QA; consumers normally use `observed_frames[]`.
- An empty `relative_path` means construct the path as
  `carla/<relative_dir>/<frame_id:06d><filename_suffix>`.
- For LiDAR, `_lidar.pcd` maps to the actual suffix `.pcd`; for Radar,
  `_radar.json` maps to `.json`.

Do not synchronize modalities by taking equal list indices or by comparing
sorted filenames. Select on the geometry clock and use the nearest observed
frame from the relevant stream.

### `carla/rsu_1/<frame_id>.json`

One raw RSU Radar frame matching `radarFrame`. It is an array of detections:

| Field | Unit | Meaning |
|---|---|---|
| `depth` | m | Range from the Radar sensor. |
| `azimuth` | rad | Horizontal detection angle in the sensor frame. |
| `altitude` | rad | Vertical detection angle in the sensor frame. |
| `velocity` | m/s | Relative radial velocity along the detection ray. |

Sensor-frame Cartesian coordinates are:

```text
x = depth * cos(altitude) * cos(azimuth)
y = depth * cos(altitude) * sin(azimuth)
z = depth * sin(altitude)
```

Use the matching per-frame RSU pose YAML when transforming these points into the
CARLA world frame.

## Read the release index and compact metadata

```python
import json
from pathlib import Path

root = Path("scene2wave_dataset")
records = [
    json.loads(line)
    for line in (root / "metadata/samples.jsonl").read_text().splitlines()
    if line.strip()
]

record = records[0]
sample = root / record["relative_path"]
metadata = json.loads((sample / "sample_metadata.json").read_text())

print(record["sample_id"])
print(metadata["radio"]["carrier_frequency_hz"])
print(metadata["channel"]["dominant_link_state"])
```

## Resolve the nearest aligned sensor frame

```python
import json
from pathlib import Path


def actual_suffix(stream):
    suffix = stream["filename_suffix"]
    if stream["modality"] == "lidar" and suffix == "_lidar.pcd":
        return ".pcd"
    if stream["modality"] == "radar" and suffix == "_radar.json":
        return ".json"
    return suffix


def nearest_stream_file(sample, alignment, stream_id, geometry_frame_id):
    stream = next(item for item in alignment["streams"] if item["stream_id"] == stream_id)
    observed = min(
        stream["observed_frames"],
        key=lambda row: abs(row["nearest_geometry_frame_id"] - geometry_frame_id),
    )
    if observed["relative_path"]:
        return sample / "carla" / observed["relative_path"]
    return (
        sample / "carla" / stream["relative_dir"]
        / f"{observed['frame_id']:06d}{actual_suffix(stream)}"
    )


alignment = json.loads((sample / "carla/alignment_index.json").read_text())
geometry_frame_id = alignment["geometry"]["frames"][1000]["frame_id"]
radar_path = nearest_stream_file(
    sample, alignment, "rsu_1.radar", geometry_frame_id
)
print(radar_path)
```

The exact Radar stream ID can also be discovered instead of assumed:

```python
radar_streams = [
    stream["stream_id"]
    for stream in alignment["streams"]
    if stream["modality"] == "radar"
]
```

## Read Radar and convert to XYZ

```python
import json
import numpy as np

detections = json.loads(radar_path.read_text())
depth = np.asarray([row["depth"] for row in detections], dtype=np.float32)
azimuth = np.asarray([row["azimuth"] for row in detections], dtype=np.float32)
altitude = np.asarray([row["altitude"] for row in detections], dtype=np.float32)
velocity = np.asarray([row["velocity"] for row in detections], dtype=np.float32)

xyz = np.column_stack(
    (
        depth * np.cos(altitude) * np.cos(azimuth),
        depth * np.cos(altitude) * np.sin(azimuth),
        depth * np.sin(altitude),
    )
)
```

## Inspect Sionna CIR/CSI NPZ

Sionna products remain in their native NPZ form. Discover keys before assuming
tensor names or dimensions:

```python
import numpy as np

npz_path = next((sample / "sionna").rglob("*_paths.npz"))
with np.load(npz_path, allow_pickle=False) as payload:
    print(npz_path.name, payload.files)
    for key in payload.files:
        print(key, payload[key].shape, payload[key].dtype)
```

Pair an NPZ frame to sensor data through the NPZ filename's frame ID and
`alignment_index.json`, not through array ordinal alone.

## Validate JSON documents

Install `jsonschema`, then select the definition appropriate to the file:

```python
import json
from jsonschema import Draft202012Validator

schema = json.loads((root / "metadata/json_schema.json").read_text())
validator = Draft202012Validator({
    "$schema": schema["$schema"],
    "$ref": "#/$defs/sampleMetadata",
    "$defs": schema["$defs"],
})
validator.validate(metadata)
```

Use `datasetInfo`, `sampleIndexRecord`, `sampleMetadata`, `alignmentIndex`, or
`radarFrame` as the `$ref` target. For JSONL, validate each line separately.