CADC

This topic describes how to manage the “CADC” dataset.

“CADC” is a fusion dataset with 8 sensors including 7 cameras and 1 lidar , and has Box3D type of labels on the point cloud data. (Fig. 8). See this page for more details about this dataset.

Fig. 8 The preview of a point cloud from “CADC” with Box3D labels.

Authorize a Client Instance

First of all, create a GAS client.

from tensorbay import GAS
from tensorbay.dataset import FusionDataset

ACCESS_KEY = "Accesskey-*****"
gas = GAS(ACCESS_KEY)

Create Fusion Dataset

Then, create a fusion dataset client by passing the fusion dataset name and is_fusion argument to the GAS client.

gas.create_dataset("CADC", is_fusion=True)

List Dataset Names

To check if you have created “CADC” fusion dataset, you can list all your available datasets. See this page for details.

The datasets listed here include both datasets and fusion datasets.

gas.list_dataset_names()

Organize Fusion Dataset

Now we describe how to organize the “CADC” fusion dataset by the FusionDataset instance before uploading it to TensorBay. It takes the following steps to organize “CADC”.

Write the Catalog

The first step is to write the catalog. Catalog is a json file contains all label information of one dataset. See this page for more details. The only annotation type for “CADC” is Box3D, and there are 10 category types and 9 attributes types.

{
    "BOX3D": {
        "isTracking": true,
        "categories": [
            { "name": "Animal" },
            { "name": "Bicycle" },
            { "name": "Bus" },
            { "name": "Car" },
            { "name": "Garbage_Container_on_Wheels" },
            { "name": "Pedestrian" },
            { "name": "Pedestrian_With_Object" },
            { "name": "Traffic_Guidance_Objects" },
            { "name": "Truck" },
            { "name": "Horse and Buggy" }
        ],
        "attributes": [
            {
                "name": "stationary",
                "type": "boolean"
            },
            {
                "name": "camera_used",
                "enum": [0, 1, 2, 3, 4, 5, 6, 7, null]
            },
            {
                "name": "state",
                "enum": ["Moving", "Parked", "Stopped"],
                "parentCategories": ["Car", "Truck", "Bus", "Bicycle", "Horse_and_Buggy"]
            },
            {
                "name": "truck_type",
                "enum": [
                    "Construction_Truck",
                    "Emergency_Truck",
                    "Garbage_Truck",
                    "Pickup_Truck",
                    "Semi_Truck",
                    "Snowplow_Truck"
                ],
                "parentCategories": ["Truck"]
            },
            {
                "name": "bus_type",
                "enum": ["Coach_Bus", "Transit_Bus", "Standard_School_Bus", "Van_School_Bus"],
                "parentCategories": ["Bus"]
            },
            {
                "name": "age",
                "enum": ["Adult", "Child"],
                "parentCategories": ["Pedestrian", "Pedestrian_With_Object"]
            },
            {
                "name": "traffic_guidance_type",
                "enum": ["Permanent", "Moveable"],
                "parentCategories": ["Traffic_Guidance_Objects"]
            },
            {
                "name": "rider_state",
                "enum": ["With_Rider", "Without_Rider"],
                "parentCategories": ["Bicycle"]
            },
            {
                "name": "points_count",
                "type": "integer",
                "minimum": 0
            }
        ]
    }
}

Note

The annotations for “CADC” have tracking information, hence the value of isTracking should be set as True.

Write the Dataloader

The second step is to write the dataloader. The dataloader function of “CADC” is to manage all the files and annotations of “CADC” into a FusionDataset instance. The code block below displays the “CADC” dataloader.

#!/usr/bin/env python3
#
# Copyright 2021 Graviti. Licensed under MIT License.
#
# pylint: disable=invalid-name
# pylint: disable=missing-module-docstring

import json
import os
from datetime import datetime
from typing import Any, Dict, List

import quaternion

from ...dataset import Data, Frame, FusionDataset
from ...exception import ModuleImportError
from ...label import LabeledBox3D
from ...sensor import Camera, Lidar, Sensors
from .._utility import glob

DATASET_NAME = "CADC"


def CADC(path: str) -> FusionDataset:
    """Dataloader of the `CADC`_ dataset.

    .. _CADC: http://cadcd.uwaterloo.ca/index.html

    The file structure should be like::

        <path>
            2018_03_06/
                0001/
                    3d_ann.json
                    labeled/
                        image_00/
                            data/
                                0000000000.png
                                0000000001.png
                                ...
                            timestamps.txt
                        ...
                        image_07/
                            data/
                            timestamps.txt
                        lidar_points/
                            data/
                            timestamps.txt
                        novatel/
                            data/
                            dataformat.txt
                            timestamps.txt
                ...
                0018/
                calib/
                    00.yaml
                    01.yaml
                    02.yaml
                    03.yaml
                    04.yaml
                    05.yaml
                    06.yaml
                    07.yaml
                    extrinsics.yaml
                    README.txt
            2018_03_07/
            2019_02_27/

    Arguments:
        path: The root directory of the dataset.

    Returns:
        Loaded `~tensorbay.dataset.dataset.FusionDataset` instance.

    """
    root_path = os.path.abspath(os.path.expanduser(path))

    dataset = FusionDataset(DATASET_NAME)
    dataset.notes.is_continuous = True
    dataset.load_catalog(os.path.join(os.path.dirname(__file__), "catalog.json"))

    for date in os.listdir(root_path):
        date_path = os.path.join(root_path, date)
        sensors = _load_sensors(os.path.join(date_path, "calib"))
        for index in os.listdir(date_path):
            if index == "calib":
                continue

            segment = dataset.create_segment(f"{date}-{index}")
            segment.sensors = sensors
            segment_path = os.path.join(root_path, date, index)
            data_path = os.path.join(segment_path, "labeled")

            with open(os.path.join(segment_path, "3d_ann.json"), "r") as fp:
                # The first line of the json file is the json body.
                annotations = json.loads(fp.readline())
            timestamps = _load_timestamps(sensors, data_path)
            for frame_index, annotation in enumerate(annotations):
                segment.append(_load_frame(sensors, data_path, frame_index, annotation, timestamps))

    return dataset


def _load_timestamps(sensors: Sensors, data_path: str) -> Dict[str, List[str]]:
    timestamps = {}
    for sensor_name in sensors.keys():
        data_folder = f"image_{sensor_name[-2:]}" if sensor_name != "LIDAR" else "lidar_points"
        timestamp_file = os.path.join(data_path, data_folder, "timestamps.txt")
        with open(timestamp_file, "r") as fp:
            timestamps[sensor_name] = fp.readlines()

    return timestamps


def _load_frame(
    sensors: Sensors,
    data_path: str,
    frame_index: int,
    annotation: Dict[str, Any],
    timestamps: Dict[str, List[str]],
) -> Frame:
    frame = Frame()
    for sensor_name in sensors.keys():
        # The data file name is a string of length 10 with each digit being a number:
        # 0000000000.jpg
        # 0000000001.bin
        data_file_name = f"{frame_index:010}"

        # Each line of the timestamps file looks like:
        # 2018-03-06 15:02:33.000000000
        timestamp = datetime.strptime(
            timestamps[sensor_name][frame_index][:23], "%Y-%m-%d %H:%M:%S.%f"
        ).timestamp()
        if sensor_name != "LIDAR":
            # The image folder corresponds to different cameras, whose name is likes "CAM00".
            # The image folder looks like "image_00".
            camera_folder = f"image_{sensor_name[-2:]}"
            image_file = f"{data_file_name}.png"

            data = Data(
                os.path.join(data_path, camera_folder, "data", image_file),
                target_remote_path=f"{camera_folder}-{image_file}",
                timestamp=timestamp,
            )
        else:
            data = Data(
                os.path.join(data_path, "lidar_points", "data", f"{data_file_name}.bin"),
                timestamp=timestamp,
            )
            data.label.box3d = _load_labels(annotation["cuboids"])

        frame[sensor_name] = data
    return frame


def _load_labels(boxes: List[Dict[str, Any]]) -> List[LabeledBox3D]:
    labels = []
    for box in boxes:
        dimension = box["dimensions"]
        position = box["position"]

        attributes = box["attributes"]
        attributes["stationary"] = box["stationary"]
        attributes["camera_used"] = box["camera_used"]
        attributes["points_count"] = box["points_count"]

        label = LabeledBox3D(
            size=(
                dimension["y"],  # The "y" dimension is the width from front to back.
                dimension["x"],  # The "x" dimension is the width from left to right.
                dimension["z"],
            ),
            translation=(
                position["x"],  # "x" axis points to the forward facing direction of the object.
                position["y"],  # "y" axis points to the left direction of the object.
                position["z"],
            ),
            rotation=quaternion.from_rotation_vector((0, 0, box["yaw"])),
            category=box["label"],
            attributes=attributes,
            instance=box["uuid"],
        )
        labels.append(label)

    return labels


def _load_sensors(calib_path: str) -> Sensors:
    try:
        import yaml  # pylint: disable=import-outside-toplevel
    except ModuleNotFoundError as error:
        raise ModuleImportError(module_name=error.name, package_name="pyyaml") from error

    sensors = Sensors()

    lidar = Lidar("LIDAR")
    lidar.set_extrinsics()
    sensors.add(lidar)

    with open(os.path.join(calib_path, "extrinsics.yaml"), "r") as fp:
        extrinsics = yaml.load(fp, Loader=yaml.FullLoader)

    for camera_calibration_file in glob(os.path.join(calib_path, "[0-9]*.yaml")):
        with open(camera_calibration_file, "r") as fp:
            camera_calibration = yaml.load(fp, Loader=yaml.FullLoader)

        # camera_calibration_file looks like:
        # /path-to-CADC/2018_03_06/calib/00.yaml
        camera_name = f"CAM{os.path.splitext(os.path.basename(camera_calibration_file))[0]}"
        camera = Camera(camera_name)
        camera.description = camera_calibration["camera_name"]

        camera.set_extrinsics(matrix=extrinsics[f"T_LIDAR_{camera_name}"])

        camera_matrix = camera_calibration["camera_matrix"]["data"]
        camera.set_camera_matrix(matrix=[camera_matrix[:3], camera_matrix[3:6], camera_matrix[6:9]])

        distortion = camera_calibration["distortion_coefficients"]["data"]
        camera.set_distortion_coefficients(**dict(zip(("k1", "k2", "p1", "p2", "k3"), distortion)))

        sensors.add(camera)
    return sensors

create a fusion dataset

To load a fusion dataset, we first need to create an instance of FusionDataset.(L75)

Note that after creating the fusion dataset, you need to set the is_continuous attribute of notes to True,(L76) since the frames in each fusion segment is time-continuous.

load the catalog

Same as dataset, you also need to load the catalog.(L77) The catalog file “catalog.json” is in the same directory with dataloader file.

create fusion segments

In this example, we create fusion segments by dataset.create_segment(SEGMENT_NAME).(L86) We manage the data under the subfolder(L33) of the date folder(L32) into a fusion segment and combine two folder names to form a segment name, which is to ensure that frames in each segment are continuous.

add sensors to fusion segments

After constructing the fusion segment, the sensors corresponding to different data should be added to the fusion segment.(L87)

In “CADC” , there is a need for projection, so we need not only the name for each sensor, but also the calibration parameters.

And to manage all the Sensors (L81, L183) corresponding to different data, the parameters from calibration files are extracted.

Lidar sensor only has extrinsics, here we regard the lidar as the origin of the point cloud 3D coordinate system, and set the extrinsics as defaults(L189).

To keep the projection relationship between sensors, we set the transform from the camera 3D coordinate system to the lidar 3D coordinate system as Camera extrinsics(L205).

Besides extrinsics(), Camera sensor also has intrinsics(), which are used to project 3D points to 2D pixels.

The intrinsics consist of two parts, CameraMatrix and DistortionCoefficients.(L208-L211)

add frames to segment

After adding the sensors to the fusion segments, the frames should be added into the continuous segment in order(L96).

Each frame contains the data corresponding to each sensor, and each data should be added to the frame under the key of sensor name(L147).

In fusion datasets, it is common that not all data have labels. In “CADC”, only point cloud files(Lidar data) have Box3D type of labels(L145). See this page for more details about Box3D annotation details.

Note

The CADC dataloader above uses relative import(L16-L19). However, when you write your own dataloader you should use regular import. And when you want to contribute your own dataloader, remember to use relative import.

Visualize Dataset

Optionally, the organized dataset can be visualized by Pharos, which is a TensorBay SDK plug-in. This step can help users to check whether the dataset is correctly organized. Please see Visualization for more details.

Upload Fusion Dataset

After you finish the dataloader and organize the “CADC” into a FusionDataset instance, you can upload it to TensorBay for sharing, reuse, etc.

# fusion_dataset is the one you initialized in "Organize Fusion Dataset" section
fusion_dataset_client = gas.upload_dataset(fusion_dataset, jobs=8)
fusion_dataset_client.commit("initial commit")

Remember to execute the commit step after uploading. If needed, you can re-upload and commit again. Please see this page for more details about version control.

Note

Commit operation can also be done on our GAS Platform.

Read Fusion Dataset

Now you can read “CADC” dataset from TensorBay.

fusion_dataset = FusionDataset("CADC", gas)

In dataset “CADC”, there are lots of FusionSegments: 2018_03_06/0001, 2018_03_07/0001, …

You can get the segment names by list them all.

fusion_dataset.keys()

You can get a segment by passing the required segment name.

fusion_segment = fusion_dataset["2018_03_06/0001"]

Note

If the segment or fusion segment is created without given name, then its name will be “”.

In the 2018_03_06/0001 fusion segment, there are several sensors. You can get all the sensors by accessing the sensors of the FusionSegment.

sensors = fusion_segment.sensors

In each fusion segment, there are a sequence of frames. You can get one by index.

frame = fusion_segment[0]

In each frame, there are several data corresponding to different sensors. You can get each data by the corresponding sensor name.

for sensor_name in sensors.keys():
    data = frame[sensor_name]

In “CADC”, only data under Lidar has a sequence of Box3D annotations. You can get one by index.

lidar_data = frame["LIDAR"]
label_box3d = lidar_data.label.box3d[0]
category = label_box3d.category
attributes = label_box3d.attributes

There is only one label type in “CADC” dataset, which is box3d. The information stored in category is one of the category names in “categories” list of catalog.json. The information stored in attributes is some of the attributes in “attributes” list of catalog.json.

See this page for more details about the structure of Box3D.

Delete Fusion Dataset

To delete “CADC”, run the following code:

gas.delete_dataset("CADC")