NN Inference

Running inference

The neural network starts at the point it is uploaded to the camera. The output from the model can be read by subscribing to a websocket endpoint ws://host:port/cameras/{mxid}/streams/{sensor_name}/nn. The outputted data will be a json in the format:

{
    "detections": ...
}

where detections will contain whetever you is your model's output.

Output format

The data format received from the websocket endpoint depends on the type of neural network you have running on the camera.

Generic neural network

In case you used a generic neural network, the output data will be a one dimensional array containing raw data from the last layer of the neural network.

{
    "detections": [
        0.3,
        0.02,
        0.05,
        ...
    ]
}

YOLO & MobilenetSSD

In case you have chosen either a YOLO or MobilenetSSD model, the output format will be an array containing detection objects. Each detection object contains:

• label - number indicating the label it detected

• confidence - float in the range [0, 1] indicating the confidence level of the detection

• coords - four element tuple, containing the coordinates of the detection in the image [x_min, y_min, x_max, y_max], these values are normalized, this means that if you want to get the real bounding box, you need to multiply both X roodinates by image width and both Y coordinates by image height

{
    "detections": [
        {"label": 0, "confidence": 0.765, "coords": [0.5, 0.1, 0.6, 0.23]},
        ...
    ]
}

Depth

In case a depth stream is active and running at the same time a neural network is deployed, each detection object will also contain information regarding its depth position (in millimeters) in addition to the information above.

{
  "detections": [
    {
      "label": 0,
      "confidence": 0.765,
      "coords": [0.5, 0.1, 0.6, 0.23],
      "spatial_coords": {
        "x": 250,
        "y": 100,
        "z": 2500
      }
    }
  ]
}