Add Smooth Loss

unsupervised/loss.py

@@ -51,3 +51,34 @@ def make_combined_ssim_l1_loss(ssim_weight: int = 0.85, other_loss_fn=l1_loss):
         return ssim_weight * ssim + (1 - ssim_weight) * other_loss_fn(target_img, reprojected_img)
 
     return combined_ssim_loss
+
+
+# TODO: Consider other gradient methods for calculating smoothness, e.g. convolution methods such as Sobel
+def smooth_loss(depth, colour_image):
+    """
+    Calculate the edge-aware per-pixel smooth loss on a depth map, with image scaled appropriately to the depth map
+
+    Does this equation (equation 3 in monodepth2 paper):
+    |dxd*t|e^(-|dxIt|) + |dyd*t|e^(-|dyIt|)
+
+    :param depth: Tensor with shape (B, h, w, 1) - disparity, such as the depth map
+    :param colour_image: Tensor with shape (B, h, w, 3) - colour image, same resolution as disparity map
+    :return: smooth loss
+    """
+    # Mean normalised inverse depth
+    normalised_depth = depth / (tf.reduce_mean(depth, [1, 2], keepdims=True) + 1e-7)
+
+    # Nothing fancy here for gradients (follows sfmlearner/monodepth), just shift 1 pixel and
+    # compare the change (x/y shift left/up 1 pixel)
+    depth_gradient_x = tf.abs(normalised_depth[:, :-1, :, :] - normalised_depth[:, 1:, :, :])
+    depth_gradient_y = tf.abs(normalised_depth[:, :, :-1, :] - normalised_depth[:, :, 1:, :])
+
+    # Colour gradients to work better with edges, monodepth 1/2 uses these
+    image_gradient_x = tf.abs(colour_image[:, :-1, :, :] - colour_image[:, 1:, :, :])
+    image_gradient_y = tf.abs(colour_image[:, :, :-1, :] - colour_image[:, :, 1:, :])
+
+    # Average the 3 colour channels into a single channel, so can be compared with the depth disparities
+    smooth_x = depth_gradient_x * tf.exp(-tf.reduce_mean(image_gradient_x, 3, keepdims=True))
+    smooth_y = depth_gradient_y * tf.exp(-tf.reduce_mean(image_gradient_y, 3, keepdims=True))
+
+    return tf.reduce_mean(smooth_x) + tf.reduce_mean(smooth_y)