Device represents an OAK camera. On all of our devices there’s a powerful vision processing unit (VPU), called Myriad X. The VPU is optimized for performing AI inference algorithms and for processing sensory inputs (eg. calculating stereo disparity from two cameras).
Device object represents an OAK device. When starting the device, you have to upload a Pipeline to it, which will get executed on the VPU.
When you create the device in the code, firmware is uploaded together with the pipeline and other assets (such as NN blobs).
pipeline = depthai.Pipeline() # Create nodes, configure them and link them together # Upload the pipeline to the device with depthai.Device(pipeline) as device: # Print Myriad X Id (MxID), USB speed, and available cameras on the device print('MxId:',device.getDeviceInfo().getMxId()) print('USB speed:',device.getUsbSpeed()) print('Connected cameras:',device.getConnectedCameras()) # Input queue, to send message from the host to the device (you can receive the message on the device with XLinkIn) input_q = device.getInputQueue("input_name", maxSize=4, blocking=False) # Output queue, to receive message on the host from the device (you can send the message on the device with XLinkOut) output_q = device.getOutputQueue("output_name", maxSize=4, blocking=False) while True: # Get a message that came from the queue output_q.get() # Or output_q.tryGet() for non-blocking # Send a message to the device cfg = depthai.ImageManipConfig() input_q.send(cfg)
Connect to specified device¶
If you have multiple devices and only want to connect to a specific one, or if your OAK PoE camera is outside of your subnet, you can specify the device (either with MxID, IP, or USB port name) you want to connect to.
# Specify MXID, IP Address or USB path device_info = depthai.DeviceInfo("14442C108144F1D000") # MXID #device_info = depthai.DeviceInfo("192.168.1.44") # IP Address #device_info = depthai.DeviceInfo("3.3.3") # USB port name with depthai.Device(pipeline, device_info) as device: # ...
After initializing the device, one has to initialize the input/output queues as well. These queues will be located on the host computer (in RAM).
outputQueue = device.getOutputQueue("output_name") inputQueue = device.getInputQueue("input_name")
When you define an output queue, the device can push new messages to it at any point in time, and the host can read from it at any point in time. Usually, when the host is reading very fast from the queue, the queue (regardless of its size) will stay empty most of the time. But as we add things on the host side (additional processing, analysis, etc), it may happen that the device will be writing to the queue faster than the host can read from it. And then the messages in the queue will start to add up - and both maxSize and blocking flags determine the behavior of the queue in this case. You can set these flags with:
# When initializing the queue queue = device.getOutputQueue(name="name", maxSize=5, blocking=False) # Or afterwards queue.setMaxSize(10) queue.setBlocking(True)
Specifying arguments for
When obtaining the output queue (example code below), the
blocking arguments should be set depending on how
the messages are intended to be used, where
name is the name of the outputting stream.
Since queues are on the host computer, memory (RAM) usually isn’t that scarce. But if you are using a small SBC like RPI Zero, where there’s only 0.5GB RAM, you might need to specify max queue size as well.
with dai.Device(pipeline) as device: queueLeft = device.getOutputQueue(name="manip_left", maxSize=8, blocking=False)
If only the latest results are relevant and previous do not matter, one can set
maxSize = 1 and
blocking = False.
That way only latest message will be kept (
maxSize = 1) and it might also be overwritten in order to avoid waiting for
the host to process every frame, thus providing only the latest data (
blocking = False).
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
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
could be set to
1, otherwise to
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.
An example would be checking the results of
DetectionNetwork for the last 1 second based on some other event,
in which case one could set
maxSize = 30 and
blocking = False
DetectionNetwork produces messages at ~30FPS).
blocking = True option is mostly used when correct order of messages is needed.
Two examples would be:
matching passthrough frames and their original frames (eg. full 4K frames and smaller preview frames that went into NN),
encoding (most prominently H264/H265 as frame drops can lead to artifacts).
By default, queues are blocking and their size is 30, so when the device fills up a queue and when the limit is
reached, any additional messages from the device will be blocked and the library will wait until it can add new messages to the queue.
It will wait for the host to consume (eg.
queue.get()) a message before putting a new one into the queue.
After the host queue gets filled up, the XLinkOut.input queue on the device will start filling up. If that queue is set to blocking, other nodes that are sending messages to it will have to wait as well. This is a usual cause for a blocked pipeline, where one of the queues isn’t emptied in timely manner and the rest of the pipeline waits for it to be empty again.
Making the queue non-blocking will change the behavior in the situation described above - instead of waiting, the library will discard
the oldest message and add the new one to the queue, and then continue its processing loop (so it won’t get blocked).
maxSize determines the size of the queue and it also helps to control memory usage.
For example, if a message has 5MB of data, and the queue size is 30, this queue can effectively store up to 150MB of data in the memory on the host (the messages can also get really big, for instance, a single 4K NV12 encoded frame takes about ~12MB).
Some additional information¶
Decreasing the queue size to 1 and setting non-blocking behavior will effectively mean “I only want the latest packet from the queue”.
Queues are thread-safe - they can be accessed from any thread.
Queues are created such that each queue is its own thread which takes care of receiving, serializing/deserializing, and sending the messages forward (same for input/output queues).
Deviceobject isn’t fully thread-safe. Some RPC calls (eg.
getDdrMemoryUsage) will get thread-safe once the mutex is set in place (right now there could be races).