VisionSensor intrinsic parameter wrong?

[Yonggie]

I printed the camera intrinsic parameters by cam.intrinsic_matrix and tried to convert depth to point cloud using the parameters. But it seem the point cloud result is dramatically incorrect? the depth image is ok, the intrinsic parameters are wrong?

I'm a bit confused in the intrinsic_matrix description

Transforming point p (x,y,z) in the camera frame via K * p will produce p' (x', y', w) - the point in the image plane. To get pixel coordiantes, divide x' and y' by w

What's K, p, w',?

my code:

# init og, cam, etc...
third_cam = VisionSensor(
        prim_path="/World/viewer_camera",
        name="my_vision_sensor",  
        modalities=["rgb", "depth"], 
        enabled=True,  
        image_height=480, 
        image_width=640,  
        focal_length=17,  
        clipping_range=(0.01, 1000000.0),  
    )
third_cam.initialize()

print(third_cam.intrinsic_matrix)
# get [[1.62252451 0.         0.        ]
# [0.         2.16336601 0.        ]
# [0.         0.         1.        ]]


# ...
def get_point_cloud(depth):
    """Projects the depth image Y into a 3D point cloud.
    Inputs:
        Y is ...xHxW   3,h,w
        camera_matrix
    Outputs:
        X is positive going right
        Y is positive into the image
        Z is positive up in the image
        XYZ is ...xHxWx3
    """
     # Extract intrinsic parameters
    K=[[1.62252451, 0.,         0.,        ],
        [0.,         2.16336601, 0.,        ],
        [0.,         0.,         1.,        ]]
    fx = K[0][0]
    fz = K[1][1]
    cx = K[0][2]
    cz = K[1][2]

    grid_x, grid_z = torch.meshgrid(torch.arange(depth.shape[-2]),
                                    torch.arange(depth.shape[-1]))

    grid_x = grid_x.unsqueeze(0).expand(depth.shape)
    grid_z = grid_z.unsqueeze(0).expand(depth.shape)

    X_t = (grid_x - cx) * depth / fx
    Z_t = (grid_z - cz) * depth / fz
    
    depth=torch.from_numpy(depth)
    XYZ = torch.stack(
        (X_t, depth, Z_t), dim=len(depth.shape))

    return XYZ

# .....

obs_dict=third_cam.get_obs()[0]
depth_image = obs_dict["depth"]
rgb_image = obs_dict["rgb"]
pc=get_point_cloud(depth_image)

[Yonggie]

The following instrinsic parameter works fine with me, with vision sensor settings being above.

K=      [[150, 0.,         80.,        ],
        [0.,         200,        60.,        ],
        [0.,         0.,         1.,        ]]