Sensor Data

LiDAR Point Cloud

LiDAR directory contains point cloud data as the name of <LIDAR_CONCAT> or <LIDAR_TOP>:

data/
└── LIDAR_CONCAT
    ├── <FRAME_ID>.pcd.bin
    ...

By default, each file contains (x, y, z, intensity, ring_idx(=-1)) as float32 values, and location coordinates are given with respect to the ego vehicle coordinate system.

Some datasets can store additional float32 features for each point, such as return_type or timestamp. In that case, sample_data.info_filename should point to a pointcloud metainfo JSON file that includes num_pts_feats, which describes the total number of float32 fields in each point.

Each file can be loaded using as follows:

# Using NumPy
import numpy as np

def load_lidar_point_cloud(file_path, num_pts_feats: int = 5) -> np.ndarray:
    data = np.fromfile(file_path, dtype=np.float32)
    return data.reshape((-1, num_pts_feats))

# Using t4-devkit
from t4_devkit.dataclass import LidarPointCloud

def load_lidar_point_cloud_t4(file_path, metainfo_path: str | None = None) -> LidarPointCloud:
    return LidarPointCloud.from_file(file_path, metainfo_filepath=metainfo_path)

When metainfo_filepath is provided and the file exists, t4-devkit uses num_pts_feats from the metainfo JSON to reshape the binary pointcloud correctly. If the metainfo file is omitted or unavailable, t4-devkit falls back to the standard 5-feature layout.

Current LidarPointCloud and SegmentationPointCloud readers keep the first 4 dimensions (x, y, z, intensity) after reshaping. Additional point features are used to determine the row width and are not exposed in the returned point matrix.

Example pointcloud metainfo JSON:

{
  "stamp": {
    "sec": 1710000000,
    "nanosec": 123456789
  },
  "num_pts_feats": 7,
  "sources": [
    {
      "sensor_token": "<SENSOR_TOKEN>",
      "idx_begin": 0,
      "length": 120000,
      "stamp": {
        "sec": 1710000000,
        "nanosec": 123456789
      }
    }
  ]
}

Camera Image

Camera directory contains raw images as the name of <CAM_XXX>:

data/
├── CAM_BACK
│   ├── <FRAME_ID>.jpg
│   ...
├── CAM_BACK_LEFT
├── CAM_BACK_RIGHT
├── CAM_FRONT
...

Radar Object

Radar directory contains radar object tracks

data/
├── RADAR_BACK
│   ├── <FRAME_ID>.pcd
│   ...
├── RADAR_BACK_LEFT
├── RADAR_BACK_RIGHT
├── RADAR_FRONT
...

Each file is based on NuScenes radar data format as follows:

# .PCD v0.7 - Point Cloud Data file format
VERSION 0.7
FIELDS x y z dyn_prop id rcs vx vy vx_comp vy_comp is_quality_valid ambig_state x_rms y_rms invalid_state pdh0 vx_rms vy_rms
SIZE 4 4 4 1 2 4 4 4 4 4 1 1 1 1 1 1 1 1
TYPE F F F I I F F F F F I I I I I I I I
COUNT 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
WIDTH 112
HEIGHT 1
VIEWPOINT 0 0 0 1 0 0 0
POINTS 112
DATA binary

from t4_devkit.dataclass import RadarPointCloud

def load_radar_point_cloud(file_path) -> RadarPointCloud:
    return RadarPointCloud.from_file(file_path)