Pipeline Issue - Input exceeds available data range

GuyGeva

Hi,

I'm trying to run a Pipeline with a MobileNetDetectionNetwork using my YOLOv5 .onnx model. I'm running into the following error:

"[14442C10815E90D000] [1.2.3.1] [1.054] [DetectionNetwork(3)] [error] Input tensor 'input' (0) exceeds available data range. Data size (143360B), tensor offset (0), size (430080B) - skipping inference"

I'm using an Oak-D device with depthai-core v2.25.1,

Here is the code running:

import blobconverter
import cv2
import depthai as dai

VIDEO_SIZE = (1072, 1072)
VIDEO_FPS = 30

pipeline = dai.Pipeline()
pipeline.setOpenVINOVersion(version=dai.OpenVINO.Version.VERSION_UNIVERSAL)
openvino_version = '2024.0.2'

def create_output(name: str, output: dai.Node.Output):
    xout = pipeline.create(dai.node.XLinkOut)
    xout.setStreamName(name)
    output.link(xout.input)

cam = pipeline.create(dai.node.ColorCamera)
cam.setPreviewSize(VIDEO_SIZE)
cam.setVideoSize(VIDEO_SIZE)
cam.setResolution(dai.ColorCameraProperties.SensorResolution.THE_1080_P)
cam.setInterleaved(False)
cam.setPreviewNumFramesPool(VIDEO_FPS)
cam.setFps(VIDEO_FPS)
cam.setBoardSocket(dai.CameraBoardSocket.CAM_A)

create_output('color', cam.video)

face_det_manip = pipeline.create(dai.node.ImageManip)
face_det_manip.initialConfig.setResize(448, 320)
face_det_manip.initialConfig.setFrameType(dai.ImgFrame.Type.GRAY8)
face_det_manip.setMaxOutputFrameSize(448*320)
face_det_manip.setNumFramesPool(VIDEO_FPS)
cam.preview.link(face_det_manip.inputImage)

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_onnx(
    model=".\\yolov5n-320x448.onnx",
    data_type="FP16",
    shaves=6
))

face_det_manip.out.link(face_det_nn.input)
create_output('detection', face_det_nn.out)

with dai.Device(pipeline) as device:
    queues = {}
    for q_name in ["color", "detection"]:
        queues[q_name] = device.getOutputQueue(q_name)

    while True:
        for q_name, q in queues.items():
            if q.has():
                msg = q.get()
                if q_name == "detection":
                    print("blah")
                if q_name == "color":
                    cv2.imshow("video", msg.getCvFrame())
                else:
                    print({q_name})

        if cv2.waitKey(1) == ord('q'):
            break

What may be causing this error?

jakaskerl

Hi @GuyGeva
Looks like the model expects RGB image, but your input is GRAY8. Change it to RGB888p.

Thanks,
Jaka

GuyGeva

Hi jakaskerl,
Changing these lines:
face_det_manip.initialConfig.setFrameType(dai.ImgFrame.Type.GRAY8) face_det_manip.setMaxOutputFrameSize(448*320)
to this:
face_det_manip.initialConfig.setFrameType(dai.ImgFrame.Type.RGB888p) face_det_manip.setMaxOutputFrameSize(448*320*3)
Causes the following error:

[14442C10815E90D000] [1.2.3.1] [1.831] [XLinkOut(4)] [error] Message has too much metadata (538047B) to serialize. Maximum is 51200B. Dropping message
[14442C10815E90D000] [1.2.3.1] [2.254] [XLinkOut(4)] [error] Message has too much metadata (537985B) to serialize. Maximum is 51200B. Dropping message
[14442C10815E90D000] [1.2.3.1] [2.528] [XLinkOut(4)] [error] Message has too much metadata (537575B) to serialize. Maximum is 51200B. Dropping message
[14442C10815E90D000] [1.2.3.1] [2.729] [XLinkOut(4)] [error] Message has too much metadata (537412B) to serialize. Maximum is 51200B. Dropping message

Is there a way to configure the maximum message size?

jakaskerl

Hi @GuyGeva
This is the Xlink from NN to host right? I think the network is producing too many detections. Can you verify the model please? Also make sure to correctly set the confidence threshold.

Thanks,
Jaka