However, if there are a lot of dropped/overwritten frames, because the host isn't able to process them fast enough

(eg. one-threaded environment which does some heavy computing), the :code:`maxSize` could be set to a higher

number, which would increase the queue size and reduce the number of dropped frames.

Specifically, at 30 FPS, a new frame is received every ~33ms, so if your host is able to process a frame in that time, the :code:`maxSize`

could be set to :code:`1`, otherwise to :code:`2` for processing times up to 66ms and so on.

If, however, there is a need to have some intervals of wait between retrieving messages, one could specify that differently.

NNData

This message carries tensors and its data. You can receive this message from the :ref:`NeuralNetwork` node or you could create this message

on the host, populate the tensor with the data and send the message to the :code:`input` of the :ref:`NeuralNetwork` node.

Reference

NeuralNetwork

This node runs neural inference on input data. Any OpenVINO neural networks can be run using this node, as long as the VPU `supports all layers <https://docs.luxonis.com/en/latest/pages/model_conversion/#unsupported-layer-type-layer-type>`__.

This allows you to pick from **200+ pre-trained model** from `Open Model Zoo <https://github.com/openvinotoolkit/open_model_zoo>`__ and `DepthAI Model Zoo <https://github.com/luxonis/depthai-model-zoo>`__

and directly run it on the OAK device.

Neural network has to be in :code:`.blob` format to be compatible with the VPU. Instructions on how to compile

your neural network (NN) to :code:`.blob` can be found `here <https://docs.luxonis.com/en/latest/pages/model_conversion/>`__.

How to place it

Object association would mean that detected objects from an external detector are mapped with tracked objects which has been detected and is being tracked from

previous frames.

Short term tracking

Short-term tracking allows to track objects between frames, thereby reducing the need to run object detection on each frame. This works great with NN

models that can't achieve 30FPS (eg. `YoloV5 <https://github.com/luxonis/depthai-experiments/tree/master/gen2-yolov5>`__); tracker can

provide tracklets when there was no inference, so the whole system can run at 30FPS.

Supported object tracker types

- :code:`SHORT_TERM_KCF`: Kernelized Correlation Filter tracking. KCF utilizes properties of circulant matrix to enhance the processing speed. Paper `here <https://www.robots.ox.ac.uk/~joao/publications/henriques_tpami2015.pdf>`__.

- :code:`SHORT_TERM_IMAGELESS`: Short-term tracking allows to track objects on frames where object detection was skipped, by extrapolating object trajectory from previous detections.

- :code:`ZERO_TERM_COLOR_HISTOGRAM`: Utilizes position, shape and input image information such as RGB histogram to perform object tracking.

- :code:`ZERO_TERM_IMAGELESS`: Only utilizes rectangular shape of detected object and position information for object tracking. It does not use color information of tracking objects. It achieves higher throughput than ZERO_TERM_COLOR_HISTOGRAM. User needs to consider the trade-off between throughput and accuracy when choosing the object tracker type.

A similar comparison of object trackers with more information can be found `here <https://github.com/openvinotoolkit/dlstreamer_gst/wiki/Object-tracking#object-tracking-types>`__.

Maximum number of tracked objects

:code:`SHORT_TERM_KCF` can track up to 60 objects at once, while all other trackers can (theoretically) track up to 1000 objects at once.

Usage

objectTracker = pipeline.create(dai.node.ObjectTracker)

objectTracker.setDetectionLabelsToTrack([15]) # Track only person

# Possible tracking types: ZERO_TERM_COLOR_HISTOGRAM, ZERO_TERM_IMAGELESS, SHORT_TERM_IMAGELESS, SHORT_TERM_KCF

objectTracker.setTrackerType(dai.TrackerType.ZERO_TERM_COLOR_HISTOGRAM)

# Take the smallest ID when new object is tracked, possible options: SMALLEST_ID, UNIQUE_ID

objectTracker.setTrackerIdAssigmentPolicy(dai.TrackerIdAssigmentPolicy.SMALLEST_ID)

auto objectTracker = pipeline.create<dai::node::ObjectTracker>();

objectTracker->setDetectionLabelsToTrack({15}); // Track only person

// Possible tracking types: ZERO_TERM_COLOR_HISTOGRAM, ZERO_TERM_IMAGELESS, SHORT_TERM_IMAGELESS, SHORT_TERM_KCF

objectTracker->setTrackerType(dai::TrackerType::ZERO_TERM_COLOR_HISTOGRAM);

// Take the smallest ID when new object is tracked, possible options: SMALLEST_ID, UNIQUE_ID

objectTracker->setTrackerIdAssigmentPolicy(dai::TrackerIdAssigmentPolicy::SMALLEST_ID);

#. `k` and `l` will decrease/increase the sensitivity iso

#. `,` and `.` will decrease/increase the focus range

.. rubric:: Similar samples:

- :ref:`Mono Camera Control`

- :ref:`Depth Crop Control`

If you want to get higher resolution RGB frames sample please visit :ref:`RGB video`.

.. rubric:: Similar samples:

- :ref:`RGB Video` (higher resolution)

- :ref:`Mono Preview`

This example shows how to use high resolution video at low latency. Compared to :ref:`RGB Preview`, this demo outputs NV12 frames whereas

preview frames are BGR and are not suited for larger resoulution (eg. 1920x1080). Preview is more suitable for either NN or visualization purposes.

.. rubric:: Similar samples:

- :ref:`RGB Preview` (lower resolution)

This example shows how to run MobileNetv2SSD on the right grayscale camera and how to display the

neural network results on a preview of the right camera stream.

.. rubric:: Similar samples:

- :ref:`RGB & MobilenetSSD`

- :ref:`RGB & MobileNetSSD @ 4K`