Is it Possible for 1 SpatialDetectionNetwork to run 2 different models?

AdamPolak

I want to do:

1. person detection (person_detection_retail_0013)

face detection (face_detection_yunet_160x120)

But I want these to be run in the "same node".

The reason is because I do not know if one, both, or neither will be detected and I want to receive synched messages by sequence number.

Would I need to make a "synchronization" node that takes in messages from 2 different SpatialDetectionNetworks and THEN send them in an xlinkout?

erik

Hi AdamPolak ,
You'd need to have 2 separate nodes, and perhaps a Script node that handles all the business logic (like syncing, or selecting which inference to run).

vital

AdamPolak Hey Adam. This works for me:

from depthai_sdk import OakCamera


synced_packets_combined = []


with OakCamera(replay='../videos/') as oak:
    color = oak.create_camera('color', )
    pdet = oak.create_nn(name='pdet', model='person-detection-retail-0013', input=color)
    fdet = oak.create_nn(name='fdet', model='face-detection-retail-0004', input=color)

    def callback(synced_packets):
        synced_packets_combined.append(synced_packets['fdet_out;color'])
        synced_packets_combined.append(synced_packets['pdet_out;color'])
        print(synced_packets_combined)

    oak.sync([pdet.out.main, fdet.out.main], callback=callback)

    oak.start(blocking=True)

This is my output (with debug level)

[184430101176431300] [2.1] [5.736] [system] [info] Memory Usage - DDR: 79.91 / 340.34 MiB, CMX: 2.48 / 2.50 MiB, LeonOS Heap: 15.55 / 77.23 MiB, LeonRT Heap: 5.52 / 41.14 MiB
[184430101176431300] [2.1] [5.736] [system] [info] Temperatures - Average: 41.78C, CSS: 43.50C, MSS 40.51C, UPA: 41.43C, DSS: 41.66C
[184430101176431300] [2.1] [5.736] [system] [info] Cpu Usage - LeonOS 26.57%, LeonRT: 55.32%
[<depthai_sdk.classes.packets.DetectionPacket object at 0x0000026B63CB9190>, <depthai_sdk.classes.packets.DetectionPacket object at 0x0000026B66B70710>, <depthai_sdk.classes.packets.DetectionPacket object at 0x0000026B63CAC550>, <depthai_sdk.classes.packets.DetectionPacket object at 0x0000026B63CAC5D0>, <depthai_sdk.classes.packets.DetectionPacket object at 0x0000026B63CADE10>, <depthai_sdk.classes.packets.DetectionPacket object at 0x0000026B63CADD90>, <depthai_sdk.classes.packets.DetectionPacket object at 0x0000026B63CAF650>, <depthai_sdk.classes.packets.DetectionPacket object at 0x0000026B63CACB10>, <depthai_sdk.classes.packets.DetectionPacket object at 0x0000026B4BBEAED0>, <depthai_sdk.classes.packets.DetectionPacket object at 0x0000026B4BBEA790>]

It keeps going...

Maybe this is useful to you?

AdamPolak

erik

I was creating a node sync script for 2 SpatialMobileDetection networks but realized I had to output to a single output or else the synchronization would be lost downstream.

Can a single XlinkOut handle 2 SpatialMobileDetection messages that are meant to be synch'd with the same sequence number? How would that work?

AdamPolak

got it thank you

erik

Hi AdamPolak ,
I'd suggest creating 2x XoutLink, so it's clear where each result came from (from which NN). I am not sure what you mean by had to output to a single output, could you elaborate?

AdamPolak

erik

I meant that I wanted a way to pick up 2 already synchronized messages on the host as a single message (or how ever I can).

But it seems like I need to go 2 XlinkOuts and then synchronize on host by checking that they have the same sequence number (I know I can make the queue size of 1 with blocking but for some reason that is not as robust?)

erik

I'd suggest doing the host-side syncing either way, where you wouldn't need to specify the queue size/blocking type if you read all new msgs fast enough.

AdamPolak

vital appreciate it thank you