Custom Decode Function

This example showcases the usage of custom decoding functions for the neural network component. More info is available inside the function itself.

Note

Visualization in current example is done with blocking behavor. This means that the program will halt at oak.start() until the window is closed. This is done to keep the example simple. For more advanced usage, see Blocking behavior section.

Setup

Please run the install script to download all required dependencies. Please note that this script must be ran from git context, so you have to download the depthai repository first and then run the script

git clone https://github.com/luxonis/depthai.git
cd depthai/
python3 install_requirements.py

For additional information, please follow our installation guide.

Pipeline

Pipeline graph

Source Code

Python

Also available on GitHub.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
import blobconverter
import numpy as np
import depthai as dai
from depthai_sdk import OakCamera
from depthai_sdk.classes import Detections


def decode(nn_data: dai.NNData) -> Detections:
    """
    Custom decode function for the NN component. Decode function has to accept NNData argument.
    The return type should preferably be a class that inherits from depthai_sdk.classes.GenericNNOutput,
    which support visualization. But this is not required, i.e. the function can return arbitrary type.

    The decoded output can be accessed from the packet object in the callback function via packet.img_detections.
    """
    layer = nn_data.getFirstLayerFp16()
    results = np.array(layer).reshape((1, 1, -1, 7))
    dets = Detections(nn_data)
    for result in results[0][0]:
        if result[2] > 0.3:
            label = int(result[1])
            conf = result[2]
            bbox = result[3:]
            det = dai.ImgDetection()
            det.confidence = conf
            det.label = label
            det.xmin = bbox[0]
            det.ymin = bbox[1]
            det.xmax = bbox[2]
            det.ymax = bbox[3]
            dets.detections.append(det)

    return dets

with OakCamera() as oak:
    color = oak.create_camera('color')

    nn_path = blobconverter.from_zoo(name='person-detection-0200', version='2021.4', shaves=6)
    nn = oak.create_nn(nn_path, color, decode_fn=decode)

    oak.visualize(nn)
    oak.start(blocking=True)

Got questions?

Head over to Discussion Forum for technical support or any other questions you might have.