Skip to content

CAD Target Training


The following requirements need to be matched in order to track a 3D model:

  • The model must be manifold
    • All surfaces need to be connected
    • No self-intersections
  • The vertex normals need to face outside of the model
  • The object model should contain less than ~10.000 faces
  • The origin of the model should be at the centroid of the object
  • The model is provided in the file format .obj (Wavefront)
  • The model dimensions in the provided file are millimeters
  • The model does not require texture information

Train CAD targets using CLI trainer

To run CAD training you need to set the

  • output directory,
  • input configuration (e.g. CADTargetTrainingConfig.json),
  • CAD file,
  • camera calibration (optional)

Detail Camera Calibration

The target data files must be generated with the camera intrinsics of the device used for tracking. Follow the instructions on Camera Calibration to generate a suitable calibration file.

Limitation: CAD Target Tracking currently only supports 640x480 input resolution (landscape mode).

Train targets for each camera (different FOV / different focal length) used for tracking.

Detail Input Configuration

The input configuration JSON file must contain the following:


  "Config.Target.Type": "CAD",
  "Config.Target.Name": "targetname",
  "CAD.UprightVector": [ 0.0, 1.0, 0.0 ],
  "CAD.Detection.DistanceRange": [ 0.3, 0.6 ],
  "CAD.Detection.NumTrainingViews": 1000,
  "CAD.Detection.ElevationAngleRange": [10.0, 70.0],
  "CAD.Model.DiscreteSymmetries": [0, 0, 0],
  "CAD.Model.ContinuousSymmetries": "",
  "CAD.Optimise": true

Configuration parameter explanation:

Always use "CAD" for CAD Training.

The name of the Target. Allowed characters: 'a-z' 'A-Z' '0-9' '_'.

A vector that defines the upright axis of the object. By default, the upright axis is defined as the Y-axis of the model, therefore [0.0, 1.0, 0.0].

Defines the detectable distance in meters. Large ranges can lead to less stable detection.

Defines the number of viewpoints for training. A higher number of viewpoints increases the detection quality. A lower number of viewpoints increases runtime performance.

The minimum and maximum elevation angle in degrees from which the object should be detectable.

Angle Description
90° Top-down view
Horizontal view
-90° Bottom-up view

Depending on the possible orientation of the target in the environment consider restricting the elevation angles to improve the detection quality.

Defines for each axis (X, Y, Z) a single rotation, in degree, representing a discrete rotational symmetry. [0, 180, 0] describes that the object is symmetric to 180 degree rotations on the Y-axis.

Defines one or multiple continuous symmetry axis. As an example, assign Y for an upright standing Cone.

Defines if the CAD model should be optimised for tracking (by default enabled). Reduces the complexity of the model.


Batch training of CAD targets is not possible.