Created
November 8, 2021 17:43
-
-
Save Joshuaalbert/725c229c85c7885b02dc32d581878536 to your computer and use it in GitHub Desktop.
Fills in empty voxels
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| def fill_in_empty_cells(voxels, length_scale_voxels=3, support=9, zero_threshold=1e-5): | |
| """ | |
| Fill in zero-values (or values less than zero_threshold) with smoothed values. | |
| Leave the non-zero bins as they are. | |
| Args: | |
| voxels: [batch, voxels_per_dimension, voxels_per_dimension, voxels_per_dimension, num_properties] | |
| support: float, length scale for exponential kernel how "near" in pixels to interpolate. | |
| support: int, how big to make the kernel, should be big enough that there are no regions of this size without a value. | |
| Returns: | |
| voxels with no zero values [batch, voxels_per_dimension, voxels_per_dimension, voxels_per_dimension, num_properties] | |
| """ | |
| # normalised filter position | |
| x = tf.range(-(support//2), (support//2)+1, 1) / length_scale_voxels | |
| X,Y,Z = tf.meshgrid(x, x, x, indexing='ij') | |
| R2 = X**2 + Y**2 + Z**2 | |
| log_filter = -0.5*R2 | |
| log_filter_sum = tf.reduce_logsumexp(log_filter) | |
| log_filter_normalised = log_filter - log_filter_sum | |
| filter = tf.math.exp(log_filter_normalised)[:,:,:,None, None]# need to be [W,H,D,1,1] | |
| # filter = tf.ones((3, 3, 3, 1, 1)) / (3. * 3. * 3.) # flat filter | |
| # [batch, voxels_per_dimension, voxels_per_dimension, voxels_per_dimension, num_properties] | |
| smoothed_voxels = tf.nn.conv3d(voxels, filters=filter, strides=[1, 1, 1, 1, 1], padding='SAME') | |
| voxels = tf.where(tf.math.abs(voxels)<zero_threshold, smoothed_voxels, voxels) | |
| return voxels |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment