Intel RealSense Depth_image problem

[Ahrovan]

Hello

Only color_images are saved. Depth_image are not saved. Please suggest a simple solution. Thank you

Related My CAR CONFIG (D435i+donkey v4.4.dev4)

# #CAMERA
CAMERA_TYPE = "D435"   # (PICAM|WEBCAM|CVCAM|CSIC|V4L|D435|MOCK|IMAGE_LIST)
IMAGE_W = 424
IMAGE_H = 240
# IMAGE_DEPTH = 3         # default RGB=3, make 1 for mono

# # Intel Realsense D435 and D435i depth sensing camera
REALSENSE_D435_RGB = True       # True to capture RGB image
REALSENSE_D435_DEPTH = True     # True to capture depth as image array
REALSENSE_D435_IMU = False      # True to capture IMU data (D435i only)
# REALSENSE_D435_ID = None        # serial number of camera or None if you only have one camera (it will autodetect)
# 

I see Depth_image clear in Code like this, but not saved

[Ezward]

@Ahrovan the depth images should be saved at 16bit png files. Are there any .png files in the images folder?

[Ahrovan]

@Ezward
no all *.jpg files (Screanshot sent above)

[Ezward]

@Ahrovan This must have gotten broken when the tub format was updated.
Yes, I just checked the code; the data type for the png is not there. I can fix this tonight.

[Ezward]

It is in version 4.1, but that saves the old tub format. I'll port the code snippet to latest at lunch and you can try on a branch.

            elif typ == 'gray16_array':
                # save np.uint16 as a 16bit png
                img = Image.fromarray(np.uint16(val))
                name = self.make_file_name(key, ext='.png')
                img.save(os.path.join(self.path, name))
                json_data[key]=name

That snipped needs to be added to tub_v2.py at around line 65, after the elif input_type == 'image_array': clause.

donkeycar/donkeycar/parts/tub_v2.py

Line 65 in c4d9091

contents[key] = name

[Ezward]

@Ahrovan I went ahead and made that change. it is is this branch 1098-fix-saving-D435-depth-image. From inside your donkeycar repo folder you can checkout the branch it that should fix the problem git checkout 1098-fix-saving-D435-depth-image Please give that a try and report back here on how it went. Thanks.

[Ahrovan]

@Ezward thank you. I will check and send result here

[Ezward]

@Ahrovan any results yet?

[Ahrovan]

@Ezward
Just modify ext to extension and for apply any filter from cv, change uint16 to uint8 (tub_v2.py)

              elif input_type == 'gray16_array':
                   # save np.uint16 as a 16bit png
                   image = Image.fromarray(np.uint8(value))
                   name = Tub._image_file_name(self.manifest.current_index, key, extension='.png')
                   image_path = os.path.join(self.images_base_path, name)
                   image.save(image_path)
                   contents[key]=name

for cv ColorMap (realsense435i.py)

            self.depth_image_r = np.asanyarray(depth_frame.get_data(), dtype=np.uint16) if self.enable_depth else None
            self.depth_image= cv2.applyColorMap(cv2.convertScaleAbs(self.depth_image_r, alpha=0.01, beta=0), cv2.COLORMAP_JET)

cv2.COLORMAP_JET can use as parameter in myConfig.py.

and for modify image size (manage.py)
need to add IMAGE_W and IMAGE_H and IMAGE_DEPTH here:

cam = RealSense435i(
            enable_rgb=cfg.REALSENSE_D435_RGB,
            enable_depth=cfg.REALSENSE_D435_DEPTH,
            enable_imu=cfg.REALSENSE_D435_IMU,
            device_id=cfg.REALSENSE_D435_ID)

[Ezward]

The issue with uint8 is that it looses accuracy, that is why I am saving it as a png. So the depth image is not meant to be a human visible image; it is meant so that you can take any pixel in the RGB image and lookup up it's depth by reading it from the depth image.

[Ahrovan]

cv2.convertScaleAbs() function returns uint8. In this case, np.uint16 gives an error. should be changed to uint8. The main problem is the need to use COLORMAP_JET or more precisely colormap. Yes, uint16 has better quality. But the problem is that the depth output does not have a suitable range, it must be corrected. Duplicate values can be seen on depth.

[Ezward]

@Ahrovan "Duplicate values can be seen on depth" What do you mean? There will certainly be some pixels that have the same depth as others; that would be normal.

To be clear, the depth image is NOT a visualization. It is data. It is the depth of each pixel in the RGB image as a 12bit value.

[Ezward]

You can see that the test program does create a visualization;

donkeycar/donkeycar/parts/realsense435i.py

Line 290 in c4d9091

depth_colormap = cv2.applyColorMap(cv2.convertScaleAbs(depth_image, alpha=0.03), cv2.COLORMAP_JET) if enable_depth else None

[Ezward]

There is an issue though; I don't save off the depth scale; without the depth scale it can't be converted back into meters. The driver does print out the depth scale when it runs, but it is not saved anywhere. A better format would just save out a 2d array of floats which represent the depth in meters for each pixel in the RGB image.

This librealsense github issue has code to show how to convert the depth data to meters. That would be easy to do. So if we want we can save both a depth visualization image and the data array as depth im meters.

# Start streaming
profile = pipeline.start(config)
# Getting the depth sensor's depth scale (see rs-align example for explanation)
depth_sensor = profile.get_device().first_depth_sensor()
depth_scale = depth_sensor.get_depth_scale()
print("Depth Scale is: " , depth_scale)
try:
    while True:

        # Wait for a coherent pair of frames: depth and color
        frames = pipeline.wait_for_frames()
        depth_frame = frames.get_depth_frame()
        color_frame = frames.get_color_frame()
        if not depth_frame or not color_frame:
            continue

        # Convert images to numpy arrays
        depth_image = np.asanyarray(depth_frame.get_data())
        color_image = np.asanyarray(color_frame.get_data())

        depth = depth_image[320,240].astype(float)
        distance = depth * depth_scale

        print ("Distance (m): ", distance )

The issue also indicates, "on all 400 Series camera models except D405 the default depth unit scale is 0.001. So if the pixel depth value was 6000 then 6000 x 0.01 would give a real-world distance of 6 meters." So we could use that value reliably. However it is probably better to just save the array a floats.

[Ezward]

For now I'm going to save the data as-is, as a 16bit grayscale png. The data can be easilly loaded by opencv or other that will load a png and the resulting image can be queried by pixel to get the distance in centimeters of that pixel. See PR #1106

[Ezward]

@Ahrovan can you confirm that the greyscale data is being saved using the branch? https://github.com/autorope/donkeycar/tree/1098-fix-saving-D435-depth-image

[Ezward]

closed as unresponsive.

[Ahrovan]

Please check this line
name = Tub._image_file_name(self.manifest.current_index, key, extension='.png')

or
name = Tub._image_file_name(self.manifest.current_index, key, ext='.png')

ext
or
extension
?
(#1106)
@Ezward