Using Custom .blob file with Oak-D lite

Andrew

I've managed to run the example python code at: https://docs.luxonis.com/software-v3/depthai/examples/detection_network/detection_network/#Detection%20network

I have a .blob file for a model that I want to run on the camera, but I can't get it to work. I've managed to get it to run but the camera feed is super slow and I'm not getting any neural network feedback. The model was trained with yolo v11. Here is the section of the code that I've modified. Thanks!

with dai.Pipeline() as pipeline:

cameraNode = pipeline.create(dai.node.Camera).build()

detectionNetwork = pipeline.create(dai.node.DetectionNetwork)

output = cameraNode.requestOutput((640, 640), dai.ImgFrame.Type.RGB888p)

output.link(detectionNetwork.input)

detectionNetwork.setBlobPath(BLOB_PATH)

labelMap = detectionNetwork.getClasses()

qRgb = detectionNetwork.passthrough.createOutputQueue()

qDet = detectionNetwork.out.createOutputQueue()

pipeline.start()

KlemenSkrlj

Hi Andrew,
could you tell me a bit more about the model? What YOLOv11 model variant are you using (e.g. nano, small, etc.)? Is it a detection only model or does it perhaps also produce instane segmentation or pose detection? Is this your first time using DepthAI and running models on the device or did perhaps had this running well before and you got this issue when transfering from DepthAIv2 to DepthAIv3?
Since you are using DAIv3 now did you perhaps consider using HubAI for conversion of the model which produces .superblob instead of .blob files allowing you to change number of shaves in the code instead of having multiple separate conversions. You can check out the docs on Hub conversion usage here.
Best,
Klemen

Andrew

Hi Klemen,

I'm not very experienced with training models, but I used roboflow and it says the type is "YOLOv11 Object Detection (accurate)." I'm pretty sure I'm only concerned with object detection at the moment. I haven't used the camera with DAIv2, but like I said I got the example model running with the code above just fine with this camera.

I did use HubAI to get the .blob file and I also saved a .superblob in case I decided to use that. Whichever one is easiest to implement is probably the one I'd prefer.

Thanks, Andrew

jasakerec

Hi @Andrew

Regarding the slow camera feed: this is expected because you use "YOLOv11 Object Detection (accurate)" which under-the-hood uses YOLOv11s (small) variant which is a little heavier than nano version. Additionally, input size 640x640 also slows down the inference so we suggest you convert the model to smaller size e.g. 320x320 - this should increase the speed. Also, consider training the model with YOLOv11 Object Detection (Fast) variant in Roboflow.

Regarding the DetectionNetwork not producing the detections: We suggest that you export your model with HubAI as you did and then download the NN archive (.tar.xz file). You dont need to extract the file, you can just give it to the DetectionNetwork with slight change in your example:

with dai.Pipeline() as pipeline:

cameraNode = pipeline.create(dai.node.Camera).build()

detectionNetwork = pipeline.create(dai.node.DetectionNetwork)

output = cameraNode.requestOutput((640, 640), dai.ImgFrame.Type.RGB888p)

output.link(detectionNetwork.input)

detectionNetwork.setNNArchive(dai.NNArchive(NN_ARCHIVE_PATH))

labelMap = detectionNetwork.getClasses()


qRgb = detectionNetwork.passthrough.createOutputQueue()

qDet = detectionNetwork.out.createOutputQueue()

pipeline.start()

The problem with your code was that DetectionNetwork can not automatically read all the parameters required for postprocessing the detections when only blob or superblob is given. The use of NN archive should fix that.

Let me know if this works for you.

Best, Jasa

Andrew

That worked, Thank you!