Hi erik

I am using Oak-1 and had depthai 2.1 installed

I now installed depthai 2.8 ...

I'm now getting rotated output without inference results.

below is the error:
[NeuralNetwork(4)] [warning] Input image (640x400) does not match NN (300x300)
FPS:0.00

Thanks & Best Regards,
Ram

  • erik replied to this.

    Hi erik

    I will wait for your response...

    Requesting your help in this regard.

    Thanks & Best Regards,
    Ram

      ramkunchur

      It looks like you'll need to use ImageManip to crop and or "squeeze" the 640x400 resolution you have to the 300x300 resolution of the neural network.

      Thoughts?

      Thanks,
      Brandon

        ramkunchur you are linking the incorrect imageManip output to the NN node. You should have linked the cropManip ImageManip node - that's the one that crops the 640x400 frames into 300x300 - to the NN input.
        Thanks, Erik

          Hi erik , Brandon

          Thanks for your reply...

          I've used same code from your example as below:

          	camRgb = pipeline.createColorCamera()
          	camRgb.setPreviewSize(640, 400)
          	camRgb.setResolution(depthai.ColorCameraProperties.SensorResolution.THE_1080_P)
          	camRgb.setInterleaved(False)
          
          	manipRgb = pipeline.createImageManip()
          	rgbRr = depthai.RotatedRect()
          	rgbRr.center.x, rgbRr.center.y = camRgb.getPreviewWidth() // 2, camRgb.getPreviewHeight() // 2
          	rgbRr.size.width, rgbRr.size.height = camRgb.getPreviewHeight(), camRgb.getPreviewWidth()
          	rgbRr.angle = 90
          	manipRgb.initialConfig.setCropRotatedRect(rgbRr, False)
          	camRgb.preview.link(manipRgb.inputImage)
          
          	cropManip = pipeline.createImageManip()
          	cropManip.initialConfig.setResize(300, 300)
          	manipRgb.out.link(cropManip.inputImage)
          
          	manipRgbOut = pipeline.createXLinkOut()
          	manipRgbOut.setStreamName("cam_out")
          	cropManip.out.link(manipRgbOut.input)

          I still get same error as below:
          [14442C1021FB92CD00] [140.524] [NeuralNetwork(4)] [warning] Input image (640x400) does not match NN (300x300)

          I do get the output frame, however, without inference results...

          Really need help to understand what I am doing wrong.
          Alternatively, can you please provide an updated script of gen2-fatigue-detection with rotate option using the code suggested for rotate?

          I am using this as an example... want to run inference with camera placed horizontally.

          This is kind of important, thanks in advance for your time and help.

          Thanks & Best Regards,
          Ram

          • erik replied to this.

            ramkunchur yes, that's the correct code. I have created another demo code that links 300x300 rotated frames to mobilenet. You will need to place this script into depthai-python/examples, as it requires mobilenet blob. Unfortunately, I don't have time to update the script you mentioned, but I am sure you will be able to update it yourself with the help of the demo script I have just created - it should be straightforward.
            Thanks, Erik

              Hi erik ...

              Thanks I'm able to get it right this time..

              However, my full screen mode doesn't work with this, probably as it needs output resolution to be in multiples of 16...

              Not sure how to resolve this as having full-screen output would have been nice

              Thanks so much for your time and help... 🙂

              Thanks & Best Regards,
              Ram

              • erik replied to this.

                ramkunchur You could just use cv2.resize() function to upscale the 300x300 frame to the desired size. You could also stream 1080P video output to the device and display detections on the video frames - not 300x300 preview frame. So something similar to this example.
                Thanks, Erik

                a year later

                Hello All,

                I am trying to rotate my camera but I am confused by the links and syntax of this api and I need this done very soon for production. Here is my code:

                def get_pipeline():
                    pipeline = dai.Pipeline()
                
                    # # Define a source - color camera
                    cam = pipeline.createColorCamera()
                    cam.setBoardSocket(dai.CameraBoardSocket.RGB)
                    # cam.setInterleaved(False)
                    cam.setResolution(dai.ColorCameraProperties.SensorResolution.THE_48_MP)
                    cam.setVideoSize(1920, 1080)
                    cam.initialControl.setSceneMode(dai.CameraControl.SceneMode.FACE_PRIORITY)
                
                    # Create MobileNet detection network
                    mobilenet = pipeline.create(dai.node.MobileNetDetectionNetwork)
                    mobilenet.setBlobPath(
                        blobconverter.from_zoo(name="face-detection-retail-0004", shaves=3)
                    )
                    mobilenet.setConfidenceThreshold(0.7)
                
                    crop_manip = pipeline.create(dai.node.ImageManip)
                    crop_manip.initialConfig.setResize(300, 300)
                    crop_manip.initialConfig.setFrameType(dai.ImgFrame.Type.BGR888p)
                    cam.isp.link(crop_manip.inputImage)
                    crop_manip.out.link(mobilenet.input)
                
                    # Create an UVC (USB Video Class) output node. It needs 1920x1080, NV12 input
                    uvc = pipeline.createUVC()
                    cam.video.link(uvc.input)

                This is what I tried but I am just guessing.

                def get_pipeline():
                pipeline = dai.Pipeline()

                    # # Define a source - color camera
                    cam = pipeline.createColorCamera()
                    cam.setBoardSocket(dai.CameraBoardSocket.RGB)
                    # cam.setInterleaved(False)
                    cam.setResolution(dai.ColorCameraProperties.SensorResolution.THE_48_MP)
                    cam.setVideoSize(1920, 1080)
                    cam.initialControl.setSceneMode(dai.CameraControl.SceneMode.FACE_PRIORITY)
                
                    # Create MobileNet detection network
                    mobilenet = pipeline.create(dai.node.MobileNetDetectionNetwork)
                    mobilenet.setBlobPath(
                        blobconverter.from_zoo(name="face-detection-retail-0004", shaves=3)
                    )
                    mobilenet.setConfidenceThreshold(0.7)
                
                    #
                
                    manipRgb = pipeline.createImageManip()
                    rgbRr = dai.RotatedRect()
                    rgbRr.center.x, rgbRr.center.y = cam.getPreviewWidth() // 2, cam.getPreviewHeight() // 2
                    rgbRr.size.width, rgbRr.size.height = cam.getPreviewHeight(), cam.getPreviewWidth()
                    rgbRr.angle = 90
                    manipRgb.initialConfig.setCropRotatedRect(rgbRr, False)
                    cam.preview.link(manipRgb.inputImage)
                
                
                    #
                
                    crop_manip = pipeline.create(dai.node.ImageManip)
                    crop_manip.initialConfig.setResize(300, 300)
                    crop_manip.initialConfig.setFrameType(dai.ImgFrame.Type.BGR888p)
                    manipRgb.out.link(crop_manip.inputImage) #added
                    cam.isp.link(crop_manip.inputImage)
                    crop_manip.out.link(mobilenet.input)
                • erik replied to this.

                  We're those guidelines for posting to the forum or submitting for review?
                  We are using a UVC and I was trying to flip the image before output but I think it needs to be 1920,1080 so it is faulting.. Is it possible to rotate the image from a script?

                  • erik replied to this.

                    Some general feedback here would be great. I do not know enough to ask the right questions yet. We have a camera using UVC and face detection but it was longer than it was tall (1920, 1080) so we wanted to rotate the and camera and stream (1080,1920). When we rotate the camera, the face detection is not looking for the sideways faces so I need to flip the stream before it goes in to that I believe but not before the UVC input? :
                    What is the max camRgb video size? We are using the OAK SOM.

                    import os
                    import sys
                    import time
                    
                    import blobconverter
                    import click
                    import depthai as dai
                    
                    if sys.version_info[0] < 3:
                        raise Exception["Doesn't work with Py2"]
                    
                    MJPEG = False
                    
                    os.environ["DEPTHAI_LEVEL"] = "debug"
                    
                    progressCalled = False
                    # TODO move this under flash(), will need to handle `progressCalled` differently
                    def progress(p):
                        global progressCalled
                        progressCalled = True
                        print(f"Flashing progress: {p*100:.1f}%")
                    
                    
                    # Will flash the bootloader if no pipeline is provided as argument
                    def flash(pipeline=None):
                        (f, bl) = dai.DeviceBootloader.getFirstAvailableDevice()
                        bootloader = dai.DeviceBootloader(bl, True)
                    
                        startTime = time.monotonic()
                        if pipeline is None:
                            print("Flashing bootloader...")
                            bootloader.flashBootloader(progress)
                        else:
                            print("Flashing application pipeline...")
                            bootloader.flash(progress, pipeline)
                    
                        if not progressCalled:
                            raise RuntimeError("Flashing failed, please try again")
                        elapsedTime = round(time.monotonic() - startTime, 2)
                        print("Done in", elapsedTime, "seconds")
                    
                    
                    @click.command()
                    @click.option(
                        "-fb",
                        "--flash-bootloader",
                        is_flag=True,
                        help="Updates device bootloader prior to running",
                    )
                    @click.option(
                        "-fp",
                        "--flash-pipeline",
                        is_flag=True,
                        help="Flashes pipeline. If bootloader flash is also requested, this will be flashed after",
                    )
                    @click.option(
                        "-gbs",
                        "--get-boot-state",
                        is_flag=True,
                        help="Prints out the boot state of the connected MX"
                    )
                    def main(flash_bootloader, flash_pipeline, get_boot_state):
                        
                        def get_pipeline():
                            pipeline = dai.Pipeline()
                    
                            # # Define a source - color camera
                            cam = pipeline.createColorCamera()
                            cam.setBoardSocket(dai.CameraBoardSocket.RGB)
                            cam.setResolution(dai.ColorCameraProperties.SensorResolution.THE_48_MP)
                            cam.setVideoSize(1920, 1080)
                            cam.initialControl.setSceneMode(dai.CameraControl.SceneMode.FACE_PRIORITY)
                    
                            # Create MobileNet detection network
                            mobilenet = pipeline.create(dai.node.MobileNetDetectionNetwork)
                            mobilenet.setBlobPath(
                                blobconverter.from_zoo(name="face-detection-retail-0004", shaves=3)
                            )
                            mobilenet.setConfidenceThreshold(0.7)
                    
                            crop_manip = pipeline.create(dai.node.ImageManip)
                            crop_manip.initialConfig.setResize(300, 300)
                            crop_manip.initialConfig.setFrameType(dai.ImgFrame.Type.BGR888p)
                            cam.isp.link(crop_manip.inputImage)
                            crop_manip.out.link(mobilenet.input)
                    
                            # Create an UVC (USB Video Class) output node. It needs 1920x1080, NV12 input
                            uvc = pipeline.createUVC()
                            cam.video.link(uvc.input)
                    
                            # Script node
                            script = pipeline.create(dai.node.Script)
                            mobilenet.out.link(script.inputs["dets"])
                            script.outputs["cam_cfg"].link(cam.inputConfig)
                            script.outputs["cam_ctrl"].link(cam.inputControl)
                            script.setScript(
                            """
                            ORIGINAL_SIZE = (5312, 6000) # 48MP with size constraints described on IMX582 luxonis page
                            SCENE_SIZE = (1920, 1080) # 1080P
                            x_arr = []
                            y_arr = []
                            AVG_MAX_NUM=7
                            limits = [0, 0] # xmin and ymin limits
                            limits.append((ORIGINAL_SIZE[0] - SCENE_SIZE[0]) / ORIGINAL_SIZE[0]) # xmax limit
                            limits.append((ORIGINAL_SIZE[1] - SCENE_SIZE[1]) / ORIGINAL_SIZE[1]) # ymax limit
                            cfg = ImageManipConfig()
                            ctrl = CameraControl()
                            def average_filter(x, y):
                                x_arr.append(x)
                                y_arr.append(y)
                                if AVG_MAX_NUM < len(x_arr): x_arr.pop(0)
                                if AVG_MAX_NUM < len(y_arr): y_arr.pop(0)
                                x_avg = 0
                                y_avg = 0
                                for i in range(len(x_arr)):
                                    x_avg += x_arr[i]
                                    y_avg += y_arr[i]
                                x_avg = x_avg / len(x_arr)
                                y_avg = y_avg / len(y_arr)
                                if x_avg < limits[0]: x_avg = limits[0]
                                if y_avg < limits[1]: y_avg = limits[1]
                                if limits[2] < x_avg: x_avg = limits[2]
                                if limits[3] < y_avg: y_avg = limits[3]
                                return x_avg, y_avg
                            while True:
                                dets = node.io['dets'].get().detections
                                if len(dets) == 0: continue
                                 coords = dets[0] # take first
                                # Get detection center
                                x = (coords.xmin + coords.xmax) / 2
                                y = (coords.ymin + coords.ymax) / 2
                                x -= SCENE_SIZE[0] / ORIGINAL_SIZE[0] / 2
                                y -= SCENE_SIZE[1] / ORIGINAL_SIZE[1] / 2
                                # node.warn(f"{x=} {y=}")
                                x_avg, y_avg = average_filter(x,y)
                                
                                # node.warn(f"{x_avg=} {y_avg=}")
                                cfg.setCropRect(x_avg, y_avg, 0, 0)
                                node.io['cam_cfg'].send(cfg)
                                node.io['cam_ctrl'].send(ctrl)
                            """
                            )
                            return pipeline
                    
                        if flash_bootloader or flash_pipeline:
                            if flash_bootloader: flash()
                            if flash_pipeline: flash(get_pipeline())
                            print("Flashing successful. Please power-cycle the device")
                            quit()
                    
                        if get_boot_state:
                            (f, bl) = dai.DeviceBootloader.getFirstAvailableDevice()
                            print(f"Device state: {bl.state.name}")
                    
                    
                        # with dai.Device(get_pipeline(), usb2Mode=True) as dev:
                        with dai.Device(get_pipeline()) as dev:
                            print(f"Connection speed: {dev.getUsbSpeed()}")
                    
                            # Doing nothing here, just keeping the host feeding the watchdog
                            while True:
                                try:
                                    time.sleep(0.1)
                                except KeyboardInterrupt:
                                    break
                    
                    
                    if __name__ == "__main__":
                        try:
                            main()
                        except KeyboardInterrupt:
                            sys.exit(0)

                    Hi chandrian ,
                    For UVC, I believe the current limitation is that frames need to be 720P and in NV12 format, so you would likely need to rotate the image after retrieving it on the host, or use some other option (eg streaming via dephtai library, then creating virtual camera on the host). Would that work for your application?
                    THanks, Erik

                      erik

                      Thanks erik you've been so helpful on this. I dont think we can flip it after the host has it... I think the idea was to rotate it so that it has more height to work with in the frame analyzing.

                      What size is the image coming out?

                      What does this crop do?:
                      crop_manip = pipeline.create(dai.node.ImageManip)
                      crop_manip.initialConfig.setResize(300, 300)
                      crop_manip.initialConfig.setFrameType(dai.ImgFrame.Type.BGR888p)
                      cam.isp.link(crop_manip.inputImage)
                      crop_manip.out.link(mobilenet.input)

                      I think the idea was to rotate it so that it has more height to work with in the frame analyzing.
                      This was the wrong assumption above. I think I can just make the face-detection-crop more tall than long and I'll be ok. It is hard to follow the dimensions.

                      Is it possible to crop into a different (smaller) size image for the face tracker? Where in the code does it need to be 1920x1080? before or after the script running?

                      • erik replied to this.

                        Hi chandrian ,

                        1. The image should be full HD if you are using depthai with UVC pipeline (docs here).
                        2. The code snippet resizes input frame to 300x300 and converts it to 8bit BGR format.
                        3. Yep that should be possible🙂
                        4. Can you please share what exactly you want to achieve?

                        Thanks, Erik