Rotate camera by 90 DEGREES

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

I am not having much success with rotating the input to the facial recognition. Do you think this is possible? If not, do you have another suggestion?

Hi chandrian , yes it should be possible.

Thanks for the reply Eric,

I have been trying to get this to work with no luck. It seems when I add multiple nodes, it does not function well. I am just feeding isp output into a mobilenet.

I tried just duplicating the resize image ImageManip above to prove functionality since I got the single resize working and I cannot get it to pass to the mobilnet and function.

This is using the UVC demo.

Thanks!

Here's the code;

#!/usr/bin/env python3

import cv2
import depthai as dai
import blobconverter

# Create pipeline
pipeline = dai.Pipeline()

# Define source and output
camRgb = pipeline.create(dai.node.ColorCamera)
xoutVideo = pipeline.create(dai.node.XLinkOut)

xoutVideo.setStreamName("video")

# Properties
camRgb.setBoardSocket(dai.CameraBoardSocket.RGB)
camRgb.setResolution(dai.ColorCameraProperties.SensorResolution.THE_1080_P)
camRgb.setVideoSize(1920, 1080)

xoutVideo.input.setBlocking(False)
xoutVideo.input.setQueueSize(1)

# 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 = camRgb.getPreviewWidth() // 2, camRgb.getPreviewHeight() // 2
# rgbRr.size.width, rgbRr.size.height = camRgb.getPreviewHeight(), camRgb.getPreviewWidth()
# rgbRr.angle = 0
# manipRgb.initialConfig.setCropRotatedRect(rgbRr, False)
#
#
# camRgb.isp.link(manipRgb.inputImage)
# manipRgb.out.link(mobilenet.input)



crop_manip2 = pipeline.create(dai.node.ImageManip)
crop_manip2.initialConfig.setResize(300, 300)
crop_manip2.initialConfig.setFrameType(dai.ImgFrame.Type.BGR888p)
camRgb.isp.link(crop_manip2.inputImage)
#crop_manip2.out.link(mobilenet.input)


crop_manip = pipeline.create(dai.node.ImageManip)
crop_manip.initialConfig.setResize(300, 300)
crop_manip.initialConfig.setFrameType(dai.ImgFrame.Type.BGR888p)
crop_manip.out.link(crop_manip2.inputImage)

# camRgb.isp.link(crop_manip.inputImage)
# crop_manip2.out.link(crop_manip.inputImage)
crop_manip.out.link(mobilenet.input)




# Script node
script = pipeline.create(dai.node.Script)
mobilenet.out.link(script.inputs["dets"])
script.outputs["cam_cfg"].link(camRgb.inputConfig)
script.outputs["cam_ctrl"].link(camRgb.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
        width = (coords.xmax - coords.xmin) * ORIGINAL_SIZE[0]
        height = (coords.ymax - coords.ymin) * ORIGINAL_SIZE[1]
        x_pixel = int(max(0, coords.xmin * ORIGINAL_SIZE[0]))
        y_pixel = int(max(0, coords.ymin * ORIGINAL_SIZE[1]))
        # ctrl.setAutoFocusRegion(x_pixel, y_pixel, int(width), int(height))
        # ctrl.setAutoExposureRegion(x_pixel, y_pixel, int(width), int(height))
        # 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)
    """
)

# Linking
camRgb.video.link(xoutVideo.input)

# Connect to device and start pipeline
with dai.Device(pipeline) as device:

    video = device.getOutputQueue(name="video", maxSize=1, blocking=False)

    while True:
        videoIn = video.get()
        print("Done in seconds")


        # Get BGR frame from NV12 encoded video frame to show with opencv
        # Visualizing the frame on slower hosts might have overhead
        cv2.imshow("video", videoIn.getCvFrame())

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

Hi chandrian ,
Please prepare the full MRE.
Thanks, Erik

erik
Hi Erik,

What do you mean when you say that? Do I post those things here in the forum post? Or do I submit something like a git issue?

Ah ok thanks. I tried to keep it minimal with the images above but maybe that was too much? Basically I fed the camera.isp output through two image crops to see if it would work and I got an error. Both image crops worked independently but it feeding through both nodes gave me an error.

Hi chandrian , I believe NNs should be run after rotating (90deg) the camera, as otherwise you will try to run inference on rotated frames (-90deg) which will have much worse performance?

I didnt quite follow. We rotated the camera so the faces coming in would be rotated 90 going into the NN so we are trying to rotate in software before the NN so that they are upright. Did I miss understand you?

I even tried it with a rotation of zero degrees to see if the facial recognition was working and it was not working.

You probably dont have time to try it yourself but if you had any guidance on the approach that would help a lot. Hoping to get this done before production next week. Thanks again

I actually got the rotated face detection working with the gen2_face detection examples so I'm trying to dissect that to make this work.

How does the "while True:" work with the script nodes? It seems like it would get stuck in there without a break..?

Hi chandrian ,
Sorry for the delay, I have been sick for the last one week. Regarding your last question; that's intentional. Otherwise the script node would be ran only once (at device startup) and never again. If you want a continuous operation, you need while True loop.
Thanks, Erik

Ok thanks for the reply. I was not able to get the rotation working for the facial recognition. I am trying to train a model for rotated faces atm.