NumPy 处理面罩形状

Question

I have an xyz point cloud ( pcd ) as a matrix of size (N, 3) and an image ( img ) as a matrix of size (H, W). I want the points where the image is projecting so I have the reunion of the following masks:

true_where_x_on_img = (0 < pcd[:, 0]) & ( pcd[:, 0] < img.shape[1])

true_where_y_on_img = (0 < pcd[:, 1]) & ( pcd[:, 1] < img.shape[0])

true_where_point_on_img = true_where_x_on_img & true_where_y_on_img

This mask is of size N and works as intended (using pcd[true_where_point_on_img]). Now I filter these values using another mask that tells me whether the pixels in the image are background or not:

true_where_not_background = mask[pcd[:, 1], pcd[:, 0]] != 0

true_where_not_background is of size M because mask is H x W. Finally, I want to project these results into column 4 of a bigger matrix, aug_pcd , of size N x 4. This matrix was initialized with zeros and pcd was copied into the first 3 columns of it. I now want to put the masked image (img[true_where_not_background]) into column 4. Something like aug_pcd[true_where_not_background, 3:] = img[true_where_not_background]. The problem is true_where_not_background has size M, and aug_pcd has row size N and is already a full-fledge slice. Slicing the slice would make a copy to which I would not be able to assign values to. How can I blend true_where_point_on_img and true_where_not_background so I can have a mask of size N?

Answer 1

SETUP

Init variables for reproducibility

pcd = np.linspace((-9, -9, -9), (10, 10, 10), 20)  # (20, 3)
img = np.random.rand(4, 8)  # (4, 8)

Mask to get points in the image

# (20,)
true_where_x_on_img = (0 < pcd[:, 0]) & (pcd[:, 0] < img.shape[1])
true_where_y_on_img = (0 < pcd[:, 1]) & (pcd[:, 1] < img.shape[0])
true_where_point_on_img = true_where_x_on_img & true_where_y_on_img

Points in image

masked_pcd = pcd[true_where_point_on_img].astype(int)  # (4, 3)

The image mask that tells the relevant pixels in the first image

# (4, 8)
img_mask = np.full(shape=(4, 8), fill_value=False, dtype=np.bool)
img_mask[1:3, 1:4] = True 

Mask for points respective to relevant pixels

# (3,)
true_where_not_zero = img_mask[masked_pcd[:, 1], masked_pcd[:, 0]] != 0

PROBLEM

The problem was that the following operation cannot be done to collect the points in the pcd that fall into the region defined by true_where_not_zero :

pcd[true_where_not_zero]
Out: Mismatched Index Error

SOLUTION

The solution was to merge the two masks into one as follows:

true_where_inside_img_and_not_zero = np.copy(true_where_point_on_img) true_where_inside_img_and_not_zero[true_where_inside_img_and_not_zero.nonzero()] = true_where_not_zero # (20,)

And pcd[true_where_inside_img_and_not_zero] would work because true_where_inside_img_and_not_zero has the same row shape as pcd