Using own blob file

GerardHarkema

Dear reader, I want to build a SpatialDetectionNetwork using the pipe below. This works well for a standard yolo_v6_nano network. However, using a custom "blob" file does not work for me.

Network code:

    def _build_and_start_pipeline(self):
        cfg = self.pipeline_config

        self.pipeline = dai.Pipeline()

        # Define sources and outputs
        self.platform = self.pipeline.getDefaultDevice().getPlatform()

        # Define sources and outputs


        size = (640, 400)

        camRgb = self.pipeline.create(dai.node.Camera).build(
            dai.CameraBoardSocket.CAM_A,
            sensorFps=cfg['fps'],
        )

        monoLeft = self.pipeline.create(dai.node.Camera).build(dai.CameraBoardSocket.CAM_B, sensorFps=cfg['fps'])
        monoRight = self.pipeline.create(dai.node.Camera).build(dai.CameraBoardSocket.CAM_C, sensorFps=cfg['fps'])
        if cfg['depth_source'] == 'stereo':
            depthSource = self.pipeline.create(dai.node.StereoDepth)
            depthSource.setExtendedDisparity(cfg['stereo_extended_disparity'])
            monoLeft.requestOutput(size).link(depthSource.left)
            monoRight.requestOutput(size).link(depthSource.right)
        elif cfg['depth_source'] == "neural":
            depthSource = self.pipeline.create(dai.node.NeuralDepth).build(
                monoLeft.requestFullResolutionOutput(),
                monoRight.requestFullResolutionOutput(),
                dai.DeviceModelZoo.NEURAL_DEPTH_LARGE,
            )
        else:
            self.get_logger().fatal(f"Unknown depth_source: {cfg['depth_source']!r}")
            raise ValueError(f"Invalid depth_source: {cfg['depth_source']}")

        test_w_yolo_v6_nano = False
        if test_w_yolo_v6_nano:
            modelDescription = dai.NNModelDescription("yolov6-nano")
            spatial_det_net = self.pipeline.create(dai.node.SpatialDetectionNetwork).build(
                camRgb, depthSource, modelDescription)
            spatial_det_net.spatialLocationCalculator.initialConfig.setSegmentationPassthrough(False)
            spatial_det_net.input.setBlocking(False)
            spatial_det_net.setDepthLowerThreshold(100)
            spatial_det_net.setDepthUpperThreshold(5000)

        else:

            spatial_det_net = self.pipeline.create(dai.node.SpatialDetectionNetwork)
            nn_size = (cfg['nn_width'], cfg['nn_height']) # 416X416 

            camRgb.requestOutput(nn_size).link(spatial_det_net.input)          
            depthSource.depth.link(spatial_det_net.inputDepth)

            spatial_det_net.setBlobPath(cfg['blob_path'])

            spatial_det_net.input.setBlocking(False)
            spatial_det_net.setConfidenceThreshold(
                float(cfg['metadata']['confidence_threshold'])
            )
            spatial_det_net.setBoundingBoxScaleFactor(0.5)
            spatial_det_net.setDepthLowerThreshold(100)
            spatial_det_net.setDepthUpperThreshold(5000)
            spatial_det_net.detectionParser.setNumClasses(int(cfg['metadata']['classes']))
            spatial_det_net.detectionParser.setCoordinateSize(
                int(cfg['metadata']['coordinates'])
            )
            spatial_det_net.detectionParser.setAnchors(
                [float(anchor) for anchor in cfg['metadata'].get('anchors', [])]
            )
            spatial_det_net.detectionParser.setAnchorMasks(
                {
                    name: [int(index) for index in indices]
                    for name, indices in cfg['metadata'].get('anchor_masks', {}).items()
                }
            )
            spatial_det_net.detectionParser.setIouThreshold(
                float(cfg['metadata']['iou_threshold'])
            )
            if cfg['labels']:
                spatial_det_net.detectionParser.setClasses(cfg['labels'])
                spatial_det_net.spatialLocationCalculator.initialConfig.setSegmentationPassthrough(
                    False
                )
            spatial_det_net.setSpatialCalculationAlgorithm(dai.SpatialLocationCalculatorAlgorithm.MIN)

        self.publisher = self.pipeline.create(Publisher)
            
        self.publisher.build(
            spatial_det_net.passthroughDepth,
            spatial_det_net.out,
            spatial_det_net.passthrough,

The output is:

[spatial_detector-2] [INFO] [1779821047.244518266] [spatial_detector]: SpatialDetector started - mode='v3' blob='SimpleFruitsYoloV5.blob' config='SimpleFruitsYoloV5.json' depth='stereo' extended_disparity=False rgb_size=416x416 fps=8.0 publish_images=True degraded_mode=False stall_timeout=5.0s image->'camera/rgb' depth->'stereo/depth' detections->'spatial_detections'
[spatial_detector-2] [INFO] [1779821047.245690628] [spatial_detector]: Pipeline thread started, calling pipeline.run()
[spatial_detector-2] [19443010611ADE1200] [1.3] [2.218] [NeuralNetwork(5)] [error] Input tensor 'images' (0) exceeds available data range. Data size (259584B), tensor offset (0), size (519168B) - skipping inference

Who can help me solving this issue. Kind regards, Gerard

OskarSonc

Hey GerardHarkema - the error suggests your custom blob is getting the wrong input frame format, not that SpatialDetectionNetwork itself is broken.

519168 = 416*416*3, while 259584 = 416*416*1.5, so your blob expects a 3-channel BGR input but the camera stream you link manually is likely NV12. The built-in yolov6-nano path works because v3 derives the required input type from model metadata automatically; with a raw .blob you need to request the camera output type explicitly.

Try changing the RGB link to:

frame_type = (
    dai.ImgFrame.Type.BGR888i
    if self.platform == dai.Platform.RVC4
    else dai.ImgFrame.Type.BGR888p
)

camRgb.requestOutput(nn_size, type=frame_type).link(spatial_det_net.input)

and keep nn_size equal to the blob input size from your JSON.

Relevant references:

Luxonis YOLO example for custom converted models: docs
Camera.requestOutput(..., type=...) docs: docs
NeuralNetwork.build() defaults to BGR888p on RVC2 and BGR888i on RVC4: GitHub

Best,
Oskar

GerardHarkema

OskarSonc

Thanks Oscar,

But now i have other messages witch i can not resolve

[spatial_detector-2] [2026-05-27 20:40:44.032] [depthai] [error] Communication exception - possible device error/misconfiguration. Original message 'Couldn't read data from stream: 'x_7passthroughDepth' (X_LINK_ERROR)'

Can you help me with these massages.

Kind regards,

Gerard

OskarSonc

GerardHarkema - those X_LINK_ERROR messages are usually not the root cause. They normally appear after the pipeline or one node has already stopped. Which device do you have?
cc @KlemenSkrlj maybe you know something that could be causing this?

Best
Oskar

GerardHarkema

Hello Oskar,

We ar using a OAK-D device.

Yo can see the complete code here: AvansMechatronica/my_depthai_ROS2blob/jazzy/my_depthai_python/my_depthai_python/spatial_detector.py

Line 390: def _build_and_start_pipeline(self)

Line 91: class Publisher(dai.node.HostNode)

Note: We use this application for students who cannot program CPP (only Python). Therefore, we do not use the ROS depthai_bridge to write ROS nodes.

Kind regards,

Gerard

KlemenSkrlj

Hey @GerardHarkema ,
Would you be able to provide a bit more information about the model? Which YOLO version is it? Where did you train it? How did you convert it?
Thanks,
Klemen

GerardHarkema

Hi Klemen,

My model is build with thes parameters:

 
 model = YOLO('yolov8n.pt')
 results = model.train(
     data=DATA_YAML,
     epochs=30,
     imgsz=640,
     batch=16,
     name='colab_yolov8_run',
     project=RESULTS_DIR,
     device=device,
 )

And it is converted by http://tools.luxonis.com/ als RVC2 model, size 640 640


You can find the model over here: https://github.com/AvansMechatronica/my_depthai_ROS2/tree/jazzy/my_depthai_python/resources

(don't look at the commit messages)


Il hope jou can give me the right solution.

Kind regards.

GerardHarkema

Hi there,

I found a solution. Converted my model to a superblob using the new Quick Conversions tool. It seems to work now fine. Thanks to Oskar & Klemen for support.

Kind regards

Gerard

KlemenSkrlj

Yeah the Hub conversion is the preferred way of converting models for DepthAIv3. This is what is actively being developed and supported. While the "tools.luxonis.com" is a deprecated path as it says on its website.
Good to hear that you good it working.
Best,
Klemen