Implement details of dense depth paper

dense_depth_functional.py

@@ -5,42 +5,37 @@ import tensorflow_datasets as tfds
 import fast_depth_functional as fd
 
 
-def dense_upproject(input, out_channels, skip_connection):
+def dense_upsample_block(input, out_channels, skip_connection):
+    """
+    Upsample block as described by dense depth in https://arxiv.org/pdf/1812.11941.pdf
+    """
     x = keras.layers.UpSampling2D(interpolation='bilinear')(input)
     x = keras.layers.Concatenate()([x, skip_connection])
     x = keras.layers.Conv2D(filters=out_channels,
                             kernel_size=3, strides=1, padding='same')(x)
-    x = keras.layers.LeakyReLU(alpha=0.2)(x)
     x = keras.layers.Conv2D(filters=out_channels,
                             kernel_size=3, strides=1, padding='same')(x)
     return keras.layers.LeakyReLU(alpha=0.2)(x)
 
 
-def dense_depth(size, weights=None, shape=(224, 224, 3), half_features=True):
+def dense_depth(size, weights=None, shape=(224, 224, 3)):
     input = keras.layers.Input(shape=shape)
     densenet = dense_net(input, size, weights, shape)
-    densenet_output_shape = densenet.layers[-1].output.shape
 
-    if half_features:
+    densenet_output_channels = densenet.layers[-1].output.shape[-1]
-        decode_filters = densenet_output_shape[-1] // 2
-    else:
-        decode_filters = int(densenet_output_shape[-1])
 
     # Reduce the feature set (pointwise)
-    decoder = keras.layers.Conv2D(filters=decode_filters, kernel_size=1, padding='same',
+    decoder = keras.layers.Conv2D(filters=densenet_output_channels, kernel_size=1, padding='same')(densenet.output)
-                                  input_shape=densenet_output_shape, name='conv2')(densenet.output)
 
     # The actual decoder
-    decoder = dense_upproject(
+    decoder = dense_upsample_block(
-        decoder, decode_filters // 2, densenet.get_layer('pool3_pool').output)
+        decoder, densenet_output_channels // 2, densenet.get_layer('pool3_pool').output)
-    decoder = dense_upproject(
+    decoder = dense_upsample_block(
-        decoder, decode_filters // 4, densenet.get_layer('pool2_pool').output)
+        decoder, densenet_output_channels // 4, densenet.get_layer('pool2_pool').output)
-    decoder = dense_upproject(
+    decoder = dense_upsample_block(
-        decoder, decode_filters // 8, densenet.get_layer('pool1').output)
+        decoder, densenet_output_channels // 8, densenet.get_layer('pool1').output)
-    decoder = dense_upproject(
+    decoder = dense_upsample_block(
-        decoder, decode_filters // 16, densenet.get_layer('conv1/relu').output)
+        decoder, densenet_output_channels // 16, densenet.get_layer('conv1/relu').output)
-    # Enable to upproject to full image size
-    # decoder = dense_upproject(decoder, int(decode_filters / 32), input)
 
     conv3 = keras.layers.Conv2D(
         filters=1, kernel_size=3, strides=1, padding='same', name='conv3')(decoder)
@@ -89,8 +84,6 @@ def dense_nnconv5(size, weights=None, shape=(224, 224, 3), half_features=True):
                    skip_connection=densenet.get_layer('pool1').output)
     x = fd.nnconv5(x, densenet.get_layer('conv1/relu').output_shape[3], 4,
                    skip_connection=densenet.get_layer('conv1/relu').output)
-    # Enable to get full dense decode (back to original size)
-    # x = fd.nnconv5(x, int(densenet.get_layer('conv1/relu').output_shape[3] / 2), 5)
 
     # Final Pointwise for depth extraction
     x = keras.layers.Conv2D(1, 1, padding='same')(x)

losses.py

@@ -2,21 +2,19 @@ import tensorflow as tf
 import tensorflow.keras.backend as K
 
 
-def dense_depth_loss_function(y_true, y_pred, theta=0.1, maxDepthVal=1000.0 / 10.0):
+def dense_depth_loss_function(y, y_pred):
-    # Point-wise depth
+    """
-    l_depth = K.mean(K.abs(y_pred - y_true), axis=-1)
+    Implementation of the loss from the dense depth paper https://arxiv.org/pdf/1812.11941.pdf
+    """
+    # Point-wise L1 loss
+    l_depth = K.mean(K.abs(y_pred - y), axis=-1)
 
-    # Edges
+    # L1 loss over image gradients
-    dy_true, dx_true = tf.image.image_gradients(y_true)
+    dy, dx = tf.image.image_gradients(y)
     dy_pred, dx_pred = tf.image.image_gradients(y_pred)
-    l_edges = K.mean(K.abs(dy_pred - dy_true) + K.abs(dx_pred - dx_true), axis=-1)
+    l_grad = K.mean(K.abs(dy_pred - dy) + K.abs(dx_pred - dx), axis=-1)
 
-    # Structural similarity (SSIM) index
+    #  Structural Similarity (SSIM)
-    l_ssim = K.clip((1 - tf.image.ssim(y_true, y_pred, maxDepthVal)) * 0.5, 0, 1)
+    l_ssim = (1 - tf.image.ssim(y, y_pred, 500)) / 2
 
-    # Weights
+    return 0.1 * K.mean(l_depth) + l_grad + l_ssim
-    w1 = 1.0
-    w2 = 1.0
-    w3 = theta
-
-    return (w1 * l_ssim) + (w2 * K.mean(l_edges)) + (w3 * K.mean(l_depth))