Python API doesn't transform correctly to point cloud from depth image

[UnaNancyOwen]

Describe the bug

Python API doesn't transform correctly to point cloud from depth image. The API returns a result, but it does not seem to be the correct point cloud.

To Reproduce

1. run this simple script for save point cloud to file

   import numpy as np
   import k4a
   
   if __name__ == '__main__':
       # Open Device
       device = k4a.Device.open()
   
       # Start Cameras
       device_config = k4a.DEVICE_CONFIG_BGRA32_1080P_NFOV_UNBINNED_FPS15
       device.start_cameras(device_config)
   
       # Get Calibration
       calibration = device.get_calibration(
           depth_mode=device_config.depth_mode,
           color_resolution=device_config.color_resolution)
   
       # Create Transformation
       transformation = k4a.Transformation(calibration)
   
       # Capture One Frame
       capture = device.get_capture(-1)
   
       # Get Point Cloud
       xyz_image = transformation.depth_image_to_point_cloud(capture.depth, k4a.ECalibrationType.DEPTH)
   
       # Save Point Cloud To Ascii Format File
       height, width, channels = xyz_image.data.shape
       xyz_data = xyz_image.data.reshape(height * width, channels) # convert to (1080 * 1920, 3) from (1080, 1920, 3)
       np.savetxt('data.txt', xyz_data, delimiter=' ', fmt='%u') # save ascii format (x y z\n x y z\n x y z\n ...)
   
       # Stop Cameras
       device.stop_cameras()

2. open ascii format point cloud data using meshlab

Expected behavior

It can get correctly transformed point cloud data.

Logs

Screenshots

Desktop (please complete the following information):

• OS with Version: Windows 10 20H2
• SDK Version: develop/HEAD
• Python Version: Python 3.8
• NumPy Version: NumPy 1.20.0

Additional context

The other features of Python API are working correctly.

[UnaNancyOwen]

@JonathanESantos What do you think? Do you have any advice?

[JonathanESantos]

There is a bug in the transformation function depth_image_to_point_cloud() where a ctypes.c_uint16 is used. It should be ctypes.c_int16. The X and Y coordinates are centered around 0 and have negative numbers.

The docstring is also a bit misleading where it says each pixel is a set of (x, y, z) values, which sounds like interleaving. In fact, the output images are arranged in planes, where the first plane is X, the next plane is Y, and the last plane is Z.

The PR to fix this is 1568. The example code to output the point cloud in a text file has also been added for reference.