def depth_to_pc(self, depth_img, pad=0.05):
"""Transforms the depth image into a point cloud representation.
Args:
depth_img (array): depth image.
Returns:
Point cloud representation of hand.
"""
# Convert to point cloud
depth_img = torch.from_numpy(depth_img).unsqueeze(0)
p_pixel = get_point_cloud(depth_img, self.num_points, 0)
p_pixel = p_pixel.squeeze(0).numpy()
# Convert to [-1, 1] cube
min_vals = p_pixel.min(0)
max_vals = p_pixel.max(0)
diff = np.abs(max_vals - min_vals)
center = (max_vals + min_vals) / 2
scale = diff[:2].max() * (0.5 + pad)
p_norm = p_pixel - center
p_norm /= scale
p_norm[:, 2] += 0.5
return p_norm
def depth_to_pc(self, depth_img, bbox, padding):
"""Transforms the depth image into a point cloud representation.
Args:
depth_img (array): depth image.
bbox (float, array): bounding box of the hand in (u, v, d).
padding (int, array): row and column padding added to the cropped
image from earlier pre-processing.
Returns:
Point cloud representation of hand.
"""
xstart = bbox[0] - padding[1][0]
ystart = bbox[2] - padding[0][0]
# Convert to point cloud
depth_img = torch.from_numpy(depth_img).unsqueeze(0)
p_ndc = get_point_cloud(depth_img, 1024, 0)
p_ndc = p_ndc.squeeze(0).numpy()
p_ndc[:, 0] += xstart
p_ndc[:, 1] += ystart
pc = self.uvd_to_xyz(p_ndc, 480, 640)
return pc