YOLO Blob Produces No Detections in DepthAI v3

yishu_corpex

Hi,

I’m trying to run a YOLO model on a OAK-D S2 PoE camera using DepthAI v3 (Python 3.1.0 bindings).
I can successfully load my YOLO .blob file and run inference with the DetectionNetwork node, but:

inDet.detections is always empty
The network outputs tensors, but no decoded bounding boxes
There is no error, just 0 detections every frame

Below is my code for you reference:

import cv2
import depthai as dai
import numpy as np

BLOB_PATH = r"C:/EMVS_Development_Camera_Scripts/Box & Label Detection/best_ckpt_openvino_2022.1_5shave.blob"
W, H = 416, 320   # model input size (width, height)

with dai.Pipeline() as pipeline:
    # Camera 
    cam = pipeline.create(dai.node.Camera).build()
    
    # Request NN-sized frames from the camera
    cam_out = cam.requestOutput(
        size=(W, H),
        type=dai.ImgFrame.Type.BGR888p,
        resizeMode=dai.ImgResizeMode.LETTERBOX,
        fps=10,
    )

    # DetectionNetwork – load local blob from disk
    det = pipeline.create(dai.node.DetectionNetwork)
    det.setBlobPath(BLOB_PATH)
    det.setConfidenceThreshold(0.5)
    det.setNumInferenceThreads(2)

    # Link camera to NN
    cam_out.link(det.input)

    # Output queues
    qDet = det.out.createOutputQueue(maxSize=4, blocking=False)
    qRgb = det.passthrough.createOutputQueue(maxSize=4, blocking=False)

    # Start pipeline
    pipeline.start()

    def frameNorm(frame, bbox):
        normVals = np.full(len(bbox), frame.shape[0])
        normVals[::2] = frame.shape[1]
        return (np.clip(np.array(bbox), 0, 1) * normVals).astype(int)

    while pipeline.isRunning():
        inFrame = qRgb.tryGet()
        inDet   = qDet.tryGet()

        if inFrame is not None:
            frame = inFrame.getCvFrame()

            # Draw detections if we also have dets
            print("InDet: ", inDet)
            if inDet is not None:
                for d in inDet.detections:
                    print("Detections Inside: ", d)
                    b = frameNorm(frame, (d.xmin, d.ymin, d.xmax, d.ymax))
                    cv2.rectangle(frame, (b[0], b[1]), (b[2], b[3]), (0, 255, 0), 2)
                    cv2.putText(frame, f"{int(d.confidence * 100)}%",
                                (b[0], b[1] - 5),
                                cv2.FONT_HERSHEY_TRIPLEX, 0.5, (0, 255, 0))

            cv2.imshow("YOLO", frame)

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

I did read here that the DetectionNetwork node replaces the YoloDetectionNetwork from v2. However, I didn't find any examples that shows how we can use a blob file of our yolo model tuned on custom data (most examples make use of yolo models from the Model Zoo) and how do we set the parameters below that were available in v2 to be set using the YoloDetectionNetwork node:

setNumClasses
setCoordinateSize
setAnchors
setAnchorMasks
setIouThreshold (for NMS algorithm)

Your help and response is highly appreciated as I am trying to resolve this issue asap.

Thanks,
Yishu

KlemenSkrlj

Hi @yishu_corpex ,
could you share a bit more information about this YOLO model (version, how you exported it - though HubAI?). And additionally could you also share a DepthAIv2 script that works for you with this exact model blob so we can check side by side that the update to DAIv3 preserved the same pipeline?
Thanks,
Klemen

yishu_corpex

Hi KlemenSkrlj

I am using Yolov6 nano and converted the pytorch model weights to blob using tools.luxonis.com. Below is the screenshot of the settings used to convert the model file to a blob. Input height is 320 and input width is 416.

I also tried converting using the HubAI platform with the settings as shown in the below screenshots. But still see the same behavior in Depthai v3 script.

Below is the depthai v2 script that works for this exact model. My aim for depthai v3 was to first have a basic script in which my model works and then upgrade it to the functionality of the below v2 script.

from pathlib import Path
import sys
import cv2
import depthai as dai
import numpy as np
import time
import argparse
import json
import blobconverter
from datetime import timedelta


# parse arguments 
parser = argparse.ArgumentParser()
parser.add_argument("-m", "--model", help="Provide model name or model path for inference",
                    default="C:/EMVS_Development_Camera_Scripts/Box & Label Detection/best_ckpt_openvino_2022.1_5shave.blob", type=str)
parser.add_argument("-c", "--config", help="Provide config path for inference",
                    default='C:/EMVS_Development_Camera_Scripts/Box & Label Detection/best_ckpt.json', type=str)
args = parser.parse_args()


# parse config
configPath = Path(args.config)
if not configPath.exists():
    raise ValueError("Path {} does not exist!".format(configPath))

with configPath.open() as f:
    config = json.load(f)
nnConfig = config.get("nn_config", {})

# parse input shape
if "input_size" in nnConfig:
    W, H = tuple(map(int, nnConfig.get("input_size").split('x')))

# extract metadata
metadata = nnConfig.get("NN_specific_metadata", {})
classes = metadata.get("classes", {})
coordinates = metadata.get("coordinates", {})
anchors = metadata.get("anchors", {})
anchorMasks = metadata.get("anchor_masks", {})
iouThreshold = metadata.get("iou_threshold", {})
confidenceThreshold = metadata.get("confidence_threshold", {})

#print(metadata)

# parse labels
nnMappings = config.get("mappings", {})
labels = nnMappings.get("labels", {})
#print("Labels: ", labels)

# get model path
nnPath = args.model
if not Path(nnPath).exists():
    print("No blob found at {}. Looking into DepthAI model zoo.".format(nnPath))
    nnPath = str(blobconverter.from_zoo(args.model, shaves = 6, zoo_type = "depthai", use_cache=True))

# sync outputs
syncNN = True


# Create pipeline
pipeline = dai.Pipeline()

# Define sources and outputs
camRgb = pipeline.create(dai.node.ColorCamera)
detectionNetwork = pipeline.create(dai.node.YoloDetectionNetwork)

# By Yishu
manip = pipeline.create(dai.node.ImageManip)

# Sync node
sync = pipeline.create(dai.node.Sync)
xoutGrp = pipeline.create(dai.node.XLinkOut)
xoutGrp.setStreamName("xout")
# Sync node properties
sync.setSyncThreshold(timedelta(milliseconds=100))

# Properties
#camRgb.setPreviewSize(W, H)

camRgb.setResolution(dai.ColorCameraProperties.SensorResolution.THE_12_MP) #THE_1080_P
camRgb.setInterleaved(False)
camRgb.setColorOrder(dai.ColorCameraProperties.ColorOrder.RGB)
camRgb.setFps(25) #40

# By Yishu
manip.initialConfig.setKeepAspectRatio(False) #True
manip.initialConfig.setResize(W, H)

# Change to RGB image than BGR - Yishu
manip.initialConfig.setFrameType(dai.ImgFrame.Type.BGR888p) #dai.ImgFrame.Type.RGB888p

# setMaxOutputFrameSize to avoid image bigger than max frame size error - Yishu
manip.setMaxOutputFrameSize(1228800)

# By Yishu
detectionNetwork.input.setQueueSize(10)

# Network specific settings
detectionNetwork.setConfidenceThreshold(confidenceThreshold)
detectionNetwork.setNumClasses(classes)
detectionNetwork.setCoordinateSize(coordinates)
detectionNetwork.setAnchors(anchors)
detectionNetwork.setAnchorMasks(anchorMasks)
detectionNetwork.setIouThreshold(iouThreshold)
detectionNetwork.setBlobPath(nnPath)
detectionNetwork.setNumInferenceThreads(2)
detectionNetwork.input.setBlocking(False)


# Linking
camRgb.isp.link(manip.inputImage)
manip.out.link(detectionNetwork.input)

# Syncing NN,ISP, Disparity, 
detectionNetwork.out.link(sync.inputs["NN_Sync"])
camRgb.isp.link(sync.inputs["ISP_Sync"])

sync.inputs["NN_Sync"].setBlocking(False)
sync.inputs["ISP_Sync"].setBlocking(False)

sync.out.link(xoutGrp.input)
xoutGrp.input.setBlocking(False)



cvColorMap = cv2.applyColorMap(np.arange(256, dtype=np.uint8), cv2.COLORMAP_JET)
cvColorMap[0] = [0, 0, 0]

# TakeUp Camera Static IP address
device_info = dai.DeviceInfo("192.168.220.10")

try:
    # Connect to device and start pipeline
    with dai.Device(pipeline, device_info) as device:
        # Output queues will be used to get the rgb frames and nn data from the outputs defined above
                        
        qSync = device.getOutputQueue("xout", 4, False)
        msgGrp = None
        syncframe = None
        sync_detections = None
        
        frame = None

        detections = []
        startTime = time.monotonic()
        counter = 0
        color2 = (255, 255, 255)

        # nn data, being the bounding box locations, are in <0..1> range - they need to be normalized with frame width/height
        def frameNorm(frame, bbox):
            normVals = np.full(len(bbox), frame.shape[0])
            normVals[::2] = frame.shape[1]
            return (np.clip(np.array(bbox), 0, 1) * normVals).astype(int)

        def displayFrame(name, frame, detections, i):
            color_box = (255, 0, 0)
            color_label = (0, 255, 0)
            color = ''
            color_blank_img = (255, 255, 255)
            blank_image = np.zeros((3040,4056,1), np.uint8)
            
            for detection in detections:
                bbox = frameNorm(frame, (detection.xmin, detection.ymin, detection.xmax, detection.ymax))
                if labels[detection.label] == 'Box':
                    color = color_box
                else:
                    color = color_label
                cv2.putText(frame, labels[detection.label], (bbox[0] + 10, bbox[1] + 20), cv2.FONT_HERSHEY_TRIPLEX, 0.5, (255,255,255))
                cv2.putText(frame, f"{int(detection.confidence * 100)}%", (bbox[0] + 10, bbox[1] + 40), cv2.FONT_HERSHEY_TRIPLEX, 0.5, (255,255,255))
                cv2.rectangle(frame, (bbox[0], bbox[1]), (bbox[2], bbox[3]), color, 2)
                cv2.rectangle(blank_image, (bbox[0], bbox[1]), (bbox[2], bbox[3]), color_blank_img, 50)
                
            cv2.namedWindow(name, cv2.WINDOW_NORMAL)
            cv2.imshow(name, frame)      


        while True:

            # By Yishu
            msgGrp = qSync.get()

            if msgGrp is not None:
                #print('msgGrp: ', msgGrp)
                for name, msg in msgGrp:
                    if name == "ISP_Sync":
                        if msg is not None:
                            syncframe = msg.getCvFrame()
                            nn_fps  =counter / (time.monotonic() - startTime)
                            print("nn_SYNC_fps: ", nn_fps)
                    
                    if name == "NN_Sync":
                        if msg is not None:
                            sync_detections = msg.detections
                            counter += 1
                    
                    if syncframe is not None and sync_detections is not None:
                        displayFrame("ISP_Sync", syncframe, sync_detections, j)
            
            if cv2.waitKey(1) == ord('q'):
                break
        
except Exception as e:
        print(f"An error occurred: {e}")
        sys.exit(0)

Thanks,
Yishu

KlemenSkrlj

One difference between the two scripts is in the resize mode: In DAIv3 you are using letterbox resizing but in DAIv2 you are squeezing/stretching since you have .setKeepAspectRation(False). Can you try using resizeMode=dai.ImgResizeMode.STRETCH or resizeMode=dai.ImgResizeMode.CROP in DAIv3 code?

yishu_corpex

KlemenSkrlj

I tried all three modes: letterbox, stretch and crop. Still i get no detections. Below is sample output message from the qdet for your reference that shows empty detections list.

sequenceNum: 73
ts:
    nsec: 157466467
    sec: 813672
tsDevice:
    nsec: 304651581
    sec: 13
detections: [ ]
transformation:
    distortionCoefficients: [3.83933, -93.8281, 0.00134366, 0.0023843, 384.297, 3.64612, -91.9485, 377.857, 0, 0, 0, 0, 0, 0,
        ]
    distortionModel: 0
    height: 320
    sourceIntrinsicMatrix: [[1025.14, 0, 672.091,
            ], [0, 1025.14, 526.316,
            ], [0, 0, 1,
            ],
        ]
    sourceIntrinsicMatrixInv: [[0.00097548, -0, -0.655612,
            ], [-0, 0.00097548, -0.513411,
            ], [0, -0, 1,
            ],
        ]
    srcCrops: [
            angle: 0
            center:
                hasNormalized: false
                normalized: false
                x: 676
                y: 506
            size:
                hasNormalized: false
                height: 1012
                normalized: false
                width: 1352,
        ]
    srcHeight: 1012
    srcWidth: 1352
    transformationMatrix: [[0.307692, 0, -1.52588e-05,
            ], [0, 0.316206, 0,
            ], [0, 0, 1,
            ],
        ]
    transformationMatrixInv: [[3.25, -0, 4.95911e-05,
            ], [-0, 3.1625, -0,
            ], [0, -0, 1,
            ],
        ]
    width: 416

Also can you let me know how can i set up these yolo config parameters in depthai v3? I have depthai 3.1 and it doesn't show me the detection parser in the detection network node. I do find the detection parser in 3.2.1 but when i use this version the entire pipeline crashes even if I am using the original code with detection parser.

setNumClasses
setCoordinateSize
setAnchors
setAnchorMasks
setIouThreshold (for NMS algorithm)

Thanks,
Yishu

jakaskerl

yishu_corpex
It's a bug in the camera node (underlying ImageManip which performs the crop).

Try using ColorCamera instead like this:

import cv2
import depthai as dai
import numpy as np

BLOB_PATH ="yolov6sr4_openvino_2022.1_5shave.blob"
W, H = 416, 320   # model input size (width, height)

with dai.Pipeline() as pipeline:
    # Camera 
    cam = pipeline.create(dai.node.ColorCamera)
    cam.setPreviewSize(W, H)
    cam.setInterleaved(False)

    cam_out = cam.preview
    
    

    # DetectionNetwork – load local blob from disk
    det = pipeline.create(dai.node.DetectionNetwork)
    det.setBlobPath(BLOB_PATH)
    det.setConfidenceThreshold(0.5)
    #det.setNumInferenceThreads(1)

    # Link camera to NN
    cam_out.link(det.input)

    # Output queues
    qDet = det.out.createOutputQueue(maxSize=4, blocking=False)
    qRgb = det.passthrough.createOutputQueue(maxSize=4, blocking=False)

    # Start pipeline
    pipeline.start()

    def frameNorm(frame, bbox):
        normVals = np.full(len(bbox), frame.shape[0])
        normVals[::2] = frame.shape[1]
        return (np.clip(np.array(bbox), 0, 1) * normVals).astype(int)

    while pipeline.isRunning():
        inFrame = qRgb.tryGet()
        inDet   = qDet.tryGet()

        if inFrame is not None:
            frame = inFrame.getCvFrame()

            # Draw detections if we also have dets
            print("InDet: ", inDet)
            if inDet is not None:
                for d in inDet.detections:
                    print("Detections Inside: ", d)
                    b = frameNorm(frame, (d.xmin, d.ymin, d.xmax, d.ymax))
                    cv2.rectangle(frame, (b[0], b[1]), (b[2], b[3]), (0, 255, 0), 2)
                    cv2.putText(frame, f"{int(d.confidence * 100)}%",
                                (b[0], b[1] - 5),
                                cv2.FONT_HERSHEY_TRIPLEX, 0.5, (0, 255, 0))

            cv2.imshow("YOLO", frame)

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

yishu_corpex

jakaskerl

I have tried using the ColorCamera node as well and the example script you have shared. I am still not seeing any detections. The detections are always empty.

When i use 3.1 the script runs but empty detections. When i use 3.2.1 the script crashes again with the error message below

DeprecationWarning: ColorCamera node is deprecated. Use Camera node instead.
  cam = pipeline.create(dai.node.ColorCamera)
InDet:  None
InDet:  None
InDet:  None
[2025-12-04 16:15:03.729] [depthai] [error] Communication exception - possible device error/misconfiguration. Original message 'Couldn't read data from stream: '__x_3_out' (X_LINK_ERROR)'
[2025-12-04 16:15:03.729] [depthai] [error] Communication exception - possible device error/misconfiguration. Original message 'Couldn't read data from stream: '__x_2_passthrough' (X_LINK_ERROR)'
[1844301021A99F0F00] [192.168.220.10] [1764882911.726] [host] [warning] Closed connection

[1844301021A99F0F00] [192.168.220.10] [1764882911.727] [host] [warning] Attempting to reconnect. Timeout is 10000ms

Thanks,
Yishu

jakaskerl

Hi yishu_corpex
Please keep using 2.31. We'll fix the issue in the meantime.

Thanks,
Jaka