Sample formats

A sample is a data point you want to label. Samples come in different types, like an image, a 3D point cloud, or a video sequence. When uploading (client.add_sample()) or downloading (client.get_sample()) a sample using the Python SDK, the format of the attributes field depends on the type of sample. The different formats are described here.

The section Import data shows how you can obtain URLs for your assets.

Image

Supported image formats: jpeg, png, bmp.

{
    "image": {
        "url": "https://example.com/image.jpg"
    }
}

If the image file is on your local computer, you should first upload it to our asset storage service (using upload_asset()) or to another cloud storage service.

Image sequence

Supported image formats: jpeg, png, bmp.

{ 
  "frames": [
    {
      "image": {
        "url": "https://example.com/frame_00001.jpg"
      },
      "name": "frame_00001" // optional
    },
    {
      "image": {
        "url": "https://example.com/frame_00002.jpg"
      },
      "name": "frame_00002"
    },
    {
      "image": {
        "url": "https://example.com/frame_00003.jpg"
      },
      "name": "frame_00003"
    }
  ]
} 

3D point cloud

On Segments.ai, the up direction is defined along the z-axis, i.e. the vector (0, 0, 1) points up. If you upload point clouds with a different up direction, you might have trouble navigating the point cloud.

{
    "pcd": {
        "url": "https://example.com/pointcloud.pcd",
        "type": "pcd"
    },
    "images": [
        { ... },
        { ... },
        { ... }
    ], // optional
    "name": "frame_00001", // optional
    "timestamp": "00001", // optional
    "ego_pose": {
        "position": {
            "x": -2.7161461413869947,
            "y": 116.25822288149078,
            "z": 1.8348751887989483
        },
        "heading": {
            "qx": -0.02111296123795955,
            "qy": -0.006495469416730261,
            "qz": -0.008024565904865688,
            "qw": 0.9997181192298087
        }
    },
    "default_z": -1, // optional, 0 by default
    "bounds": { // optional
        "min_z": -1,
        "max_z": 3
    }
}

Name	Type	Description
pcd	Point cloud data	Required. Point cloud data.
images	array of camera images	Reference camera images.
name	string	Name of the sample.
timestamp	int, float, or string	Timestamp of the sample. Should be in nanoseconds for accurate velocity/acceleration calculations. Will also be used for interpolation unless disabled in dataset settings.
ego_pose	Ego pose	Pose of the sensor that captured the point cloud data.
default_z	float	Default z-value of the ground plane. 0 by default. Only valid in the point cloud cuboid editor. New cuboids will be drawn on top of the ground plane, i.e. the default z-position of a new cuboid is 0.5 (since the default height of a new cuboid is 1).
bounds	dict	Point cloud bounds: a dict with values min_z and max_z (float). These values are used as the default min/max values for height coloring when provided.

Point cloud data

See 3D point cloud formats for the supported file formats.

{
    "url": "https://example.com/pointcloud.bin",
    "type": "kitti"
}

Name	Type	Description
url	string	Required. URL of the point cloud data.
type	string: "pcd" | "binary-xyzi" | "kitti" | "binary-xyzir" | "nuscenes" | "ply"	Required. Type of the point cloud data. See #3d-point-cloud file formats for the list of supported file formats.

If the point cloud file is on your local computer, you should first upload it to our asset storage service (using upload_asset()) or to another cloud storage service.

Camera image

A calibrated or uncalibrated reference image corresponding to a point cloud. The reference images can be opened in a new tab from within the labeling interface. You can determine the layout of the images by setting the row and col attributes on each image. If you also supply the calibration parameters (and distortion parameters if necessary), the main point cloud view can be set to the image to obtain a fused view.

{
    "name": "Camera example 1", // optional
    "url": "https://example.com/image.jpg",
    "row": 0,
    "col": 0,
    "intrinsics": { // optional
        "intrinsic_matrix": [
            [1266.417203046554, 0, 816.2670197447984],
            [0, 1266.417203046554, 491.50706579294757],
            [0, 0, 1]
        ]
    },
    "extrinsics": { // optional
        "translation": {
            "x": -0.012463384576629082,
            "y": 0.76486688894964,
            "z": -0.3109103442096661
        },
        "rotation": {
            "qx": 0.713640516187247,
            "qy": -0.001134052598226082,
            "qz": 0.0036449450274057696,
            "qw": 0.7005017073187271
        }
    },
    "distortion": { // optional
        "model": "fisheye",
        "coefficients": {
            "k1": -0.0539124,
            "k2": -0.0101993,
            "k3": -0.00202017,
            "k4": 0.00120938
        }
    },
    "camera_convention": "OpenCV", // optional
    "rotation": 1.5708 // optional
}

Name	Type	Description
name	string	Name of the camera image.
url	string	Required. URL of the camera image.
row	int	Required. Row of this image in the images viewer.
col	int	Required. Column of this image in the images viewer.
intrinsics	Camera intrinsics	Intrinsic parameters of the camera.
extrinsics	Camera extrinsics	Extrinsic parameters of the camera relative to the ego pose.
distortion	Distortion	Distortion parameters of the camera.
camera_convention	string: "OpenGL" | "OpenCV"	Convention of the camera coordinates. We use the OpenGL/Blender coordinate convention for cameras. +X is right, +Y is up, and +Z is pointing back and away from the camera. -Z is the look-at direction. Other codebases may use the OpenCV convention, where the Y and Z axes are flipped but the +X axis remains the same. See diagram 1.
rotation	float	The rotation that needs to be applied when displaying the image. Valid options are 0, $\frac{\pi}{4}$, $\frac{\pi}{2}$, and $\frac{3 \pi}{4}$. Useful for when a camera is mounted upside-down.

If the image file is on your local computer, you should first upload it to our asset storage service (using upload_asset()) or to another cloud storage service.

Camera intrinsics

{
    "intrinsic_matrix": [
        [1266.417203046554, 0, 816.2670197447984],
        [0, 1266.417203046554, 491.50706579294757],
        [0, 0, 1]
    ]
}

Name	Type	Description
intrinsic_matrix	2D array of floats representing 3x3 matrix $K$in row-major order​	Required. Intrinsic matrix $K$ used in the pinhole camera model. $K = \begin{bmatrix} f_x & 0 & o_x\\ 0 & f_y & o_y \\ 0 & 0 & 1 \end{bmatrix}$ ​$f_x$ and $f_y$​ are the focal lengths in pixels. We assume square pixels, so $f_x = f_y$​. $o_x$ and $o_y$ are the offsets (in pixels) of the principal point from the top-left corner of the image frame.

Camera extrinsics

{
    "translation": {
        "x": -0.012463384576629082,
        "y": 0.76486688894964,
        "z": -0.3109103442096661
    },
    "rotation": {
        "qx": 0.713640516187247,
        "qy": -0.001134052598226082,
        "qz": 0.0036449450274057696,
        "qw": 0.7005017073187271
    }
}

Name	Type	Description
translation	object: { "x": float, "y": float, "z": float }	Required. Translation of the camera in lidar coordinates, i.e., relative to the ego pose.
rotation	object: { "qx": float, "qy": float, "qz": float, "qw": float }	Required. Rotation of the camera in lidar coordinates, i.e., relative to the ego pose (or equivalently: a transformation from camera frame to ego frame). Defined as a rotation quaternion. By default, we use the OpenGL/Blender coordinate convention for cameras. +X is right, +Y is up, and +Z is pointing back and away from the camera. -Z is the look-at direction. Other codebases may use the OpenCV convention, where the Y and Z axes are flipped but the +X axis remains the same. See diagram 1. You can specify the camera convention in Camera image.

Diagram 1: camera convention for calibrated camera images on Segments.ai.

Distortion

// Fisheye
{ 
    "model": "fisheye",
    "coefficients": {
        "k1": -0.0539124,
        "k2": -0.0101993,
        "k3": -0.00202017,
        "k4": 0.00120938
}
// Brown-Conrady
{ 
    "model": "brown-conrady",
    "coefficients": {
        "k1": -0.2916058942,
        "k2": 0.0763231072,
        "k3": 0.0,
        "p1": 0.0014829263,
        "p2": -0.0019540316
    }
}

Name	Type	Description
model	string: "fisheye" | "brown-conrady"	Required. Type of the distortion model: fisheye or Brown-Conrady.
coefficients	Fisheye: object: { "k1": float, "k2": float, "k3": float, "k4": float, } Brown-Conrady: object: { "k1": float, "k2": float, "k3": float, "p1": float, "p2": float }	Required. Coefficients of the distortion model: k1, k2, k3, k4 for fisheye (see the OpenCV fisheye model) and k1, k2, k3, p1, p2 for Brown-Conrady (see the OpenCV distortion model, note that $k_4$ and $k_5$ are not used).

Ego pose

The pose of the sensor used to capture the 3D point cloud data. This can be helpful if you want to obtain cuboids in world coordinates, or when your sensor is moving. In the latter situation, supplying an ego pose with each frame will ensure that static objects do not move when switching between frames.

{
    "position": {
        "x": -2.7161461413869947,
        "y": 116.25822288149078,
        "z": 1.8348751887989483
    },
    "heading": {
        "qx": -0.02111296123795955,
        "qy": -0.006495469416730261,
        "qz": -0.008024565904865688,
        "qw": 0.9997181192298087
    }
},

Name	Type	Description
position	object: { "x": float, "y": float, "z": float }	Required. XYZ position of the sensor in world coordinates.
heading	object: { "qx": float, "qy": float, "qz": float, "qw": float }	Required. Orientation of the sensor. Defined as a rotation quaternion.

Segments.ai uses 32-bit floats for the point positions. Keep in mind that 32-bit floats have limited precision. In fact, only 24 bits can be used to represent the number itself (the significand, excluding the sign bit), or about 7.22 decimal digits. If you want to keep two decimal places, this only leaves 5.22 decimal digits, so the numbers shouldn't be larger than 10^5.22 = 165958.

To avoid rounding problems, it is best practice to subtract the ego position of the first frame from all other ego positions. This way, the first ego position is set to (0, 0, 0) and the subsequent ego positions are relative to (0, 0, 0) . In your export script, you can add the ego position of the first frame back to the object positions.

3D point cloud sequence

{ 
  "frames": [
    { ... },
    { ... },
    { ... }
  ]
} 

Name	Type	Description
frames	array of 3D point clouds	Required. List of 3D point cloud frames in the sequence.

Multi-sensor sequence

{
  "sensors": [
    {
      "name": "Lidar", 
      "task_type": "pointcloud-cuboid-sequence",
      "attributes": { ... }
    },
    {
      "name": "Camera 1", 
      "task_type": "image-vector-sequence",
      "attributes": { ... } 
    },
    ...
  ]
}

Name	Type	Description
sensors	array of sensors	Required. List of the sensors that can be labeled.

Sensor

Name	Type	Description
name	string	Required. The name of the sensor.
task_type	string	Required. The task type of the sensor. Currently, pointcloud-cuboid-sequence and image-vector-sequence are supported.
attributes	object	Required. The sample attributes for the sensor. Currently, 3D point cloud sequence and image sequence are supported.