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


      Thanks again for the response Erik. Basically we need to zoom in on the person like this and crop to this more vertical size.

      • erik replied to this.

        Hi chandrian ,
        with UVC mode this (currently) isn't possible, as UVC node needs full hd images. You could, however, stream exact same image but rotated by 90deg. Thoughts?
        Thanks, Erik

        Yes I attempted that but was not successful. Can you give me general instructions of where to implement that? The problems I faced were that the UVC needed 1920x1080 and when I rotated that, it was 1080x1920, and that the face recognition did not work when the camera was rotated 90 degrees.

        Thanks,
        Aaron

        • erik replied to this.

          Hi chandrian ,
          I assume you are using something similar to Lossless Zooming. So first you would want to rotate the frame 90deg (so people are upright), do the face detection, crop the original (rotated) 4k image into 1080x1920 (as in the lossless zooming example), then rotate that to 1080P, which you can feed into the UVC node. Thougths?
          Thanks ,Erik

          Ok so this wouldnt be in the script then. I realize script is mostly for changing the pipeline anyway. Yes that sounds like a plan for me. I will attempt and let you know. Thanks!!

          I will probably need to remove this before the rotate then? : cam.setVideoSize(1920, 1080)

          • erik replied to this.

            Hi chandrian , by default you will want to rotate the images by 90deg. So you will likely want 4k, then rotate it by 90deg, then do inference, then crop, then rotate back by -90deg to get to 1920x1080.

            Ok thanks! Is all of this happening in before the script node? Or is that unnecessary.

            And how does the script node work in terms of code path. I see a "while true" in the script with no breaks and a while true after the script. do they run in parallel?

            I tried keeping the same dimensions as my working code and just flipping twice and I didnt not get an output stream and then I tried a zero degree turn twice and still no stream. Am I messing something up here:

                    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 = 0
                    manipRgb.initialConfig.setCropRotatedRect(rgbRr, False)
                    cam.preview.link(manipRgb.inputImage)
            
                    manipRgb2 = pipeline.createImageManip()
                    manipRgb2.initialConfig.setCropRotatedRect(rgbRr, False)
                    manipRgb.out.link(manipRgb2.inputImage)
            
                    # Create an UVC (USB Video Class) output node. It needs 1920x1080, NV12 input
                    uvc = pipeline.createUVC()
                    manipRgb2.out.link(uvc.input)

            I actually cant get the cam.video to go through any manipulation node and into the UVC

            I tried passing the cam.video into the manip node and into the uvc. Then I tried setting the preview to 1920x1080 (is that a possible size?) and feeding that into the manip node and into uvc and I still could not get that working either.

            • erik replied to this.

              Hi chandrian ,
              With the new depthai you can also use cam.video with ImageManip. I believe we plan to update the depthai uvc branch to latest, so you will be able to achieve this. Regarding the issue, please submit the full MRE.
              Thanks, Erik

              Ok I will try to submit that. I have a deadline soon so I am not sure that will be done in time. Do you think it would be possible to rotate the facial recognition input so that, if the camera is 90 rotated, it will still recognize faces? I will try that today but no luck so far. Actually I think its working now.. more details to come
              Thanks,
              Aaron

              edit:
              Facial recognition seems to be working (blue square coming up) but not tracking at this moment.
              edit2:
              I think the blue squares were windows camera app tracking face, not the depthai.