[NeuralNetwork(7)] [error] Input tensor 'data' (0) exceeds available data range. Data size (3600B), tensor offset (0), size (10800B) - skipping inference

[vipulkumar-developer]

Hi, I'm trying yo manipulate a black and white frame obtained from the Right Monocamera using the Script node.

In the script file I try these lines of code:

cfg = ImageManipConfig()
cfg.setFrameType(RawImgFrame.Type.RGB888p)

But the problem is that the pipeline seems to ignore it since the resulting error is the following: [14442C10310D57D700] [483.907] [NeuralNetwork(7)] [error] Input tensor 'data' (0) exceeds available data range. Data size (3600B), tensor offset (0), size (10800B) - skipping inference

[Erol444]

Hi @vipulkumar-developer , I am not sure why you need Script node, but if you want to run NN that expects 3 channels (RGB) on mono frames, please see line 42 of this example: https://docs.luxonis.com/projects/api/en/latest/samples/MobileNet/mono_mobilenet/#mono-mobilenetssd Thanks, Erik

[vipulkumar-developer]

Yes, it is exactly what I'm trying to do but in the script file for the Script node. I'm working with the Script node since I'm changing an existent repository.

[Erol444]

@vipulkumar-developer I see, could you prepare a full MRE, please?

[Erol444]

@vipulkumar-developer please see the documentation on how to prepare a MRE. Thanks, Erik

[vipulkumar-developer]

My objective is to use the repository gen2-face-recognition but working with black and white frames, not RGB.

In the main.py file I've modified the pipeline by inserting the MonoCamera and in some components adding initialConfig.setFrameType(dai.RawImgFrame.Type.RGB888p):

def create_pipeline():
    pipeline = dai.Pipeline()
    openvino_version = '2021.2'

    cam = pipeline.create(dai.node.MonoCamera)
    cam.setResolution(dai.MonoCameraProperties.SensorResolution.THE_400_P)
    cam.setBoardSocket(dai.CameraBoardSocket.LEFT)

    host_face_out = pipeline.create(dai.node.XLinkOut)
    host_face_out.setStreamName('frame')
    cam.out.link(host_face_out.input)

    # NeuralNetwork
    print("Creating Face Detection Neural Network...")
    face_det_nn = pipeline.create(dai.node.MobileNetDetectionNetwork)
    face_det_nn.setConfidenceThreshold(0.5)
    face_det_nn.setBlobPath(blobconverter.from_zoo(
        name="face-detection-retail-0004",
        shaves=6,
        version=openvino_version
    ))

    # Link Face ImageManip -> Face detection NN node
    face_det_manip.out.link(face_det_nn.input)

    # Script node will take the output from the face detection NN as an input and set ImageManipConfig
    # to the 'age_gender_manip' to crop the initial frame
    script = pipeline.create(dai.node.Script)
    script.setProcessor(dai.ProcessorType.LEON_CSS)
    with open("script.py", "r") as f:
        script.setScript(f.read())

    face_det_nn.out.link(script.inputs['face_det_in'])
    # We are only interested in timestamp, so we can sync depth frames with NN output
    face_det_nn.passthrough.link(script.inputs['face_pass'])

    # ImageManip as a workaround to have more frames in the pool.
    # cam.preview can only have 4 frames in the pool before it will
    # wait (freeze). Copying frames and setting ImageManip pool size to
    # higher number will fix this issue.
    copy_manip = pipeline.create(dai.node.ImageManip)
    cam.out.link(copy_manip.inputImage)
    copy_manip.setNumFramesPool(20)
    copy_manip.setMaxOutputFrameSize(720*1280*3)

    copy_manip.out.link(face_det_manip.inputImage)
    copy_manip.out.link(script.inputs['preview'])

    print("Creating Head pose estimation NN")
    headpose_manip = pipeline.create(dai.node.ImageManip)
    headpose_manip.initialConfig.setResize(60, 60)
    headpose_manip.initialConfig.setFrameType(dai.RawImgFrame.Type.RGB888p)

    script.outputs['manip_cfg'].link(headpose_manip.inputConfig)
    script.outputs['manip_img'].link(headpose_manip.inputImage)

    return pipeline

While in the scriptcopy.py file I've added initialConfig.setFrameType(dai.RawImgFrame.Type.RGB888p) in the following piece of code:

        node.warn("Entering in the for")
        for det in face_dets.detections:
            bboxes.append(det) # For the rotation
            cfg = ImageManipConfig()
            correct_bb(det)
            cfg.setCropRect(det.xmin, det.ymin, det.xmax, det.ymax)
            cfg.setResize(60, 60)
            cfg.setFrameType(RawImgFrame.Type.RGB888p)
            cfg.setKeepAspectRatio(False)
            node.io['manip_cfg'].send(cfg)
            node.io['manip_img'].send(img)

The blob NN models used are the same of the original repo.

When I start the script I obtain this error: [14442C10310D57D700] [483.907] [NeuralNetwork(7)] [error] Input tensor 'data' (0) exceeds available data range. Data size (3600B), tensor offset (0), size (10800B) - skipping inference

I suppose that is related to some ImageManipNode that resizes the frame to 60x60 (3600), but it expects an RGB frame since 3600 * 3 = 10800

[Erol444]

@vipulkumar-developer please see again the MRE tutorialm, and provide everything in single zip file, and reduce the script - headpose estimation has nothing to to with the issue you are having.

[vipulkumar-developer]

Here is the zip