Cuboid annotation format

Description of cuboid annotations exported in the Mindkosh format

Each annotation item is of the form:

{
    "id": <annotation_id - unique for every instance of an object>,
    "track_id": <object id - unique for every object, null if it isnt a track>,
    "objectType": <classname>,
    "h": <height>,
    "w": <width>,
    "l": <length>,
    "keyframe": <Is this frame manually labeled? Applicable only for tracks>
    "item": {
        "tx": <x-co-ordinate of the center>,
        "ty": <y-co-ordinate of the center>,
        "tz": <z-co-ordinate of the center>,
        "rx": <rotation in radians around X-axis>,
        "ry": <rotation in radians around Y-axis>,
        "rz": <rotation in radians around Z-axis>,
        "occlusion": 0,
        "face": <FACE_ID describing which way the object is facing>
    }
    "attributes": [
        {
            "spec_id": <attribute_id>,
            "value": <value of the attribute>
        }
    ]
}

Face ID

FACE_ID is encoded as follows. Note that the angle of rotation alone is enough to deduce the front face. FACE_ID is a redundant field to deduce the same.

ID	Direction vector
0	+X
1	-X
2	+Y
3	-Y
4	+Z
5	-Z

Conventions

We follow the same convention as KITTI, with 1 major change - all our values are specified from the point cloud frame, rather than the camera frame.

Dimensions

1. Length is measured front-to-back.

2. Width is measured side-to-side

3. Height is measured from top-to-bottom

This assumes that the forward facing faces have been properly marked. By default the face is assumed to be in the positive Y direction (when the cuboid is un-rotated).

Rotation

Rotation is measured counter-clockwise from the Y axis (of the point cloud). It is always positive.

An example scenario when looking at an object top-down. Red axis is X, Green axis is Y.