Correct computing offset of anchors.

Check #1073 for more
information.

keras_retinanet/backend/common.py

@@ -52,16 +52,22 @@ def bbox_transform_inv(boxes, deltas, mean=None, std=None):
     return pred_boxes
 
 
-def shift(shape, stride, anchors):
+def shift(image_shape, features_shape, stride, anchors):
     """ Produce shifted anchors based on shape of the map and stride size.
 
     Args
-        shape  : Shape to shift the anchors over.
-        stride : Stride to shift the anchors with over the shape.
-        anchors: The anchors to apply at each location.
+        image_shape    : Shape of the input image.
+        features_shape : Shape of the feature map.
+        stride         : Stride to shift the anchors with over the image.
+        anchors        : The anchors to apply at each location.
     """
-    shift_x = (keras.backend.arange(0, shape[1], dtype=keras.backend.floatx()) + keras.backend.constant(0.5, dtype=keras.backend.floatx())) * stride
-    shift_y = (keras.backend.arange(0, shape[0], dtype=keras.backend.floatx()) + keras.backend.constant(0.5, dtype=keras.backend.floatx())) * stride
+    # compute the offset of the anchors based on the image shape and the feature map shape
+    # see https://github.com/fizyr/keras-retinanet/issues/1073 for more information
+    offset_x = keras.backend.cast((image_shape[1] - (features_shape[1] - 1) * stride), keras.backend.floatx()) / 2.0
+    offset_y = keras.backend.cast((image_shape[0] - (features_shape[0] - 1) * stride), keras.backend.floatx()) / 2.0
+
+    shift_x = keras.backend.arange(0, features_shape[1], dtype=keras.backend.floatx()) * stride + offset_x
+    shift_y = keras.backend.arange(0, features_shape[0], dtype=keras.backend.floatx()) * stride + offset_y
 
     shift_x, shift_y = meshgrid(shift_x, shift_y)
     shift_x = keras.backend.reshape(shift_x, [-1])

keras_retinanet/layers/_misc.py

@@ -58,28 +58,29 @@ def __init__(self, size, stride, ratios=None, scales=None, *args, **kwargs):
         super(Anchors, self).__init__(*args, **kwargs)
 
     def call(self, inputs, **kwargs):
-        features = inputs
-        features_shape = keras.backend.shape(features)
+        image, features = inputs
+        features_shape  = keras.backend.shape(features)
+        image_shape     = keras.backend.shape(image)
 
         # generate proposals from bbox deltas and shifted anchors
         if keras.backend.image_data_format() == 'channels_first':
-            anchors = backend.shift(features_shape[2:4], self.stride, self.anchors)
+            anchors = backend.shift(image_shape[2:4], features_shape[2:4], self.stride, self.anchors)
         else:
-            anchors = backend.shift(features_shape[1:3], self.stride, self.anchors)
+            anchors = backend.shift(image_shape[1:3], features_shape[1:3], self.stride, self.anchors)
         anchors = keras.backend.tile(keras.backend.expand_dims(anchors, axis=0), (features_shape[0], 1, 1))
 
         return anchors
 
     def compute_output_shape(self, input_shape):
-        if None not in input_shape[1:]:
+        if None not in input_shape[1][1:]:
             if keras.backend.image_data_format() == 'channels_first':
-                total = np.prod(input_shape[2:4]) * self.num_anchors
+                total = np.prod(input_shape[1][2:4]) * self.num_anchors
             else:
-                total = np.prod(input_shape[1:3]) * self.num_anchors
+                total = np.prod(input_shape[1][1:3]) * self.num_anchors
 
-            return (input_shape[0], total, 4)
+            return (input_shape[1][0], total, 4)
         else:
-            return (input_shape[0], None, 4)
+            return (input_shape[1][0], None, 4)
 
     def get_config(self):
         config = super(Anchors, self).get_config()

keras_retinanet/models/retinanet.py

@@ -207,11 +207,12 @@ def __build_pyramid(models, features):
     return [__build_model_pyramid(n, m, features) for n, m in models]
 
 
-def __build_anchors(anchor_parameters, features):
+def __build_anchors(anchor_parameters, image, features):
     """ Builds anchors for the shape of the features from FPN.
 
     Args
         anchor_parameters : Parameteres that determine how anchors are generated.
+        image             : The image input tensor.
         features          : The FPN features.
 
     Returns
@@ -229,7 +230,7 @@ def __build_anchors(anchor_parameters, features):
             ratios=anchor_parameters.ratios,
             scales=anchor_parameters.scales,
             name='anchors_{}'.format(i)
-        )(f) for i, f in enumerate(features)
+        )([image, f]) for i, f in enumerate(features)
     ]
 
     return keras.layers.Concatenate(axis=1, name='anchors')(anchors)
@@ -328,7 +329,7 @@ def retinanet_bbox(
 
     # compute the anchors
     features = [model.get_layer(p_name).output for p_name in ['P3', 'P4', 'P5', 'P6', 'P7']]
-    anchors  = __build_anchors(anchor_params, features)
+    anchors  = __build_anchors(anchor_params, model.inputs[0], features)
 
     # we expect the anchors, regression and classification values as first output
     regression     = model.outputs[0]

keras_retinanet/utils/anchors.py

@@ -234,24 +234,29 @@ def anchors_for_shape(
             ratios=anchor_params.ratios,
             scales=anchor_params.scales
         )
-        shifted_anchors = shift(image_shapes[idx], anchor_params.strides[idx], anchors)
+        shifted_anchors = shift(image_shape, image_shapes[idx], anchor_params.strides[idx], anchors)
         all_anchors     = np.append(all_anchors, shifted_anchors, axis=0)
 
     return all_anchors
 
 
-def shift(shape, stride, anchors):
-    """ Produce shifted anchors based on shape of the map and stride size.
+def shift(image_shape, features_shape, stride, anchors):
+    """ Produce shifted anchors based on shape of the image, shape of the feature map and stride.
 
     Args
-        shape  : Shape to shift the anchors over.
-        stride : Stride to shift the anchors with over the shape.
-        anchors: The anchors to apply at each location.
+        image_shape    : Shape of the input image.
+        features_shape : Shape of the feature map.
+        stride         : Stride to shift the anchors with over the image.
+        anchors        : The anchors to apply at each location.
     """
+    # compute the offset of the anchors based on the image shape and the feature map shape
+    # see https://github.com/fizyr/keras-retinanet/issues/1073 for more information
+    offset_x = (image_shape[1] - (features_shape[1] - 1) * stride) / 2.0
+    offset_y = (image_shape[0] - (features_shape[0] - 1) * stride) / 2.0
 
     # create a grid starting from half stride from the top left corner
-    shift_x = (np.arange(0, shape[1]) + 0.5) * stride
-    shift_y = (np.arange(0, shape[0]) + 0.5) * stride
+    shift_x = np.arange(0, features_shape[1]) * stride + offset_x
+    shift_y = np.arange(0, features_shape[0]) * stride + offset_y
 
     shift_x, shift_y = np.meshgrid(shift_x, shift_y)
 

tests/layers/test_misc.py

@@ -30,12 +30,16 @@ def test_simple(self):
             scales=np.array([1], keras.backend.floatx()),
         )
 
+        # create fake image input (only shape is used anyway)
+        image = np.zeros((1, 16, 16, 3), dtype=keras.backend.floatx())
+        image = keras.backend.variable(image)
+
         # create fake features input (only shape is used anyway)
         features = np.zeros((1, 2, 2, 1024), dtype=keras.backend.floatx())
         features = keras.backend.variable(features)
 
         # call the Anchors layer
-        anchors = anchors_layer.call(features)
+        anchors = anchors_layer.call([image, features])
         anchors = keras.backend.eval(anchors)
 
         # expected anchor values
@@ -59,12 +63,16 @@ def test_mini_batch(self):
             scales=np.array([1], dtype=keras.backend.floatx()),
         )
 
+        # create fake image input (only shape is used anyway)
+        image = np.zeros((2, 16, 16, 3), dtype=keras.backend.floatx())
+        image = keras.backend.variable(image)
+
         # create fake features input with batch_size=2
         features = np.zeros((2, 2, 2, 1024), dtype=keras.backend.floatx())
         features = keras.backend.variable(features)
 
         # call the Anchors layer
-        anchors = anchors_layer.call(features)
+        anchors = anchors_layer.call([image, features])
         anchors = keras.backend.eval(anchors)
 
         # expected anchor values

tests/utils/test_anchors.py

@@ -167,3 +167,29 @@ def test_anchors_for_shape_values():
         strides[0] * 3 / 2 + (sizes[0] * scales[1] / np.sqrt(ratios[1])) / 2,
         strides[0] * 3 / 2 + (sizes[0] * scales[1] * np.sqrt(ratios[1])) / 2,
     ], decimal=6)
+
+
+def test_anchors_for_shape_odd_input():
+    pyramid_levels = [3]
+    image_shape    = (20, 20)  # this shape causes rounding errors when downsampling using convolutions
+    sizes          = [32]
+    strides        = [8]
+    ratios         = np.array([1], keras.backend.floatx())
+    scales         = np.array([1], keras.backend.floatx())
+    anchor_params  = AnchorParameters(sizes, strides, ratios, scales)
+
+    anchors = anchors_for_shape(image_shape, pyramid_levels = pyramid_levels, anchor_params = anchor_params)
+
+    expected_anchors = np.array([
+        [-14, -14, 18, 18],
+        [-6 , -14, 26, 18],
+        [2  , -14, 34, 18],
+        [-14, -6 , 18, 26],
+        [-6 , -6 , 26, 26],
+        [2  , -6 , 34, 26],
+        [-14,  2 , 18, 34],
+        [-6 ,  2 , 26, 34],
+        [2  ,  2 , 34, 34],
+    ])
+
+    np.testing.assert_equal(anchors, expected_anchors)