maximum proposals

keras_rcnn/backend/tensorflow_backend.py

@@ -56,7 +56,7 @@ def non_maximum_suppression(boxes, scores, maximum, threshold=0.5):
     )
 
 
-def propose(boxes, scores):
+def propose(boxes, scores, maximum):
     shape = keras.backend.int_shape(boxes)[1:3]
 
     shifted = keras_rcnn.backend.shift(shape, 16)
@@ -79,7 +79,7 @@ def propose(boxes, scores):
     proposals = keras.backend.cast(proposals, tensorflow.float32)
     scores = keras.backend.cast(scores, tensorflow.float32)
 
-    indicies = keras_rcnn.backend.non_maximum_suppression(proposals, scores, 100, 0.7)
+    indicies = keras_rcnn.backend.non_maximum_suppression(proposals, scores, maximum, 0.7)
 
     proposals = keras.backend.gather(proposals, indicies)
 

keras_rcnn/layers/object_detection/_object_proposal.py

@@ -19,7 +19,7 @@ def build(self, input_shape):
         super(ObjectProposal, self).build(input_shape)
 
     def call(self, inputs, **kwargs):
-        return keras_rcnn.backend.propose(inputs[0], inputs[1])
+        return keras_rcnn.backend.propose(inputs[0], inputs[1], self.proposals)
 
     def compute_output_shape(self, input_shape):
         return self.output_dim

tests/conftest.py

@@ -14,6 +14,49 @@ def anchor_layer():
     return keras_rcnn.layers.object_detection.Anchor(features, shape)
 
 
+@pytest.fixture()
+def convolution_neural_network():
+    options = {
+        "activation": "relu",
+        "kernel_size": (3, 3),
+        "padding": "same"
+    }
+
+    shape = (224, 224, 3)
+
+    x = keras.layers.Input(shape)
+
+    y = keras.layers.Conv2D(64, **options)(x)
+    y = keras.layers.Conv2D(64, **options)(y)
+
+    y = keras.layers.MaxPooling2D(strides=(2, 2))(y)
+
+    y = keras.layers.Conv2D(128, **options)(y)
+    y = keras.layers.Conv2D(128, **options)(y)
+
+    y = keras.layers.MaxPooling2D(strides=(2, 2))(y)
+
+    y = keras.layers.Conv2D(256, **options)(y)
+    y = keras.layers.Conv2D(256, **options)(y)
+    y = keras.layers.Conv2D(256, **options)(y)
+    y = keras.layers.Conv2D(256, **options)(y)
+
+    y = keras.layers.MaxPooling2D(strides=(2, 2))(y)
+
+    y = keras.layers.Conv2D(512, **options)(y)
+    y = keras.layers.Conv2D(512, **options)(y)
+    y = keras.layers.Conv2D(512, **options)(y)
+    y = keras.layers.Conv2D(512, **options)(y)
+
+    y = keras.layers.MaxPooling2D(strides=(2, 2))(y)
+
+    y = keras.layers.Conv2D(512, **options)(y)
+    y = keras.layers.Conv2D(512, **options)(y)
+    y = keras.layers.Conv2D(512, **options)(y)
+
+    return keras.models.Model(x, y)
+
+
 @pytest.fixture()
 def feat_h():
     return 14
@@ -72,6 +115,59 @@ def regional_proposal_network_layer():
     return keras_rcnn.layers.object_detection.RegionProposalNetwork()
 
 
+@pytest.fixture()
+def region_proposal_network():
+    options = {
+        "activation": "relu",
+        "kernel_size": (3, 3),
+        "padding": "same"
+    }
+
+    shape = (224, 224, 3)
+
+    x = keras.layers.Input(shape)
+
+    y = keras.layers.Conv2D(64, **options)(x)
+    y = keras.layers.Conv2D(64, **options)(y)
+
+    y = keras.layers.MaxPooling2D(strides=(2, 2))(y)
+
+    y = keras.layers.Conv2D(128, **options)(y)
+    y = keras.layers.Conv2D(128, **options)(y)
+
+    y = keras.layers.MaxPooling2D(strides=(2, 2))(y)
+
+    y = keras.layers.Conv2D(256, **options)(y)
+    y = keras.layers.Conv2D(256, **options)(y)
+    y = keras.layers.Conv2D(256, **options)(y)
+    y = keras.layers.Conv2D(256, **options)(y)
+
+    y = keras.layers.MaxPooling2D(strides=(2, 2))(y)
+
+    y = keras.layers.Conv2D(512, **options)(y)
+    y = keras.layers.Conv2D(512, **options)(y)
+    y = keras.layers.Conv2D(512, **options)(y)
+    y = keras.layers.Conv2D(512, **options)(y)
+
+    y = keras.layers.MaxPooling2D(strides=(2, 2))(y)
+
+    y = keras.layers.Conv2D(512, **options)(y)
+    y = keras.layers.Conv2D(512, **options)(y)
+    y = keras.layers.Conv2D(512, **options)(y)
+
+    y = keras.layers.Conv2D(512, **options)(y)
+
+    a = keras.layers.Conv2D(9 * 1, (1, 1), activation="sigmoid")(y)
+
+    b = keras.layers.Conv2D(9 * 4, (1, 1))(y)
+
+    y = keras_rcnn.layers.object_detection.ObjectProposal(300)([a, b])
+
+    model = keras.models.Model(x, y)
+
+    model.compile("sgd", "mse")
+
+
 @pytest.fixture()
 def x():
     return numpy.arange(2 * 9 * 14 * 14, dtype=numpy.float32).reshape(1, 18, 14, 14)

tests/layers/object_detection/test_object_proposal.py

@@ -1,19 +1,70 @@
+import keras.layers
+import keras.models
 import numpy
 
+import keras_rcnn.layers.object_detection
+
 
 class TestObjectProposal:
     def test_build(self):
         assert True
 
-    def test_call(self, object_proposal_model):
-        object_proposal_model.compile("sgd", "mse")
+    def test_call(self):
+        options = {
+            "activation": "relu",
+            "kernel_size": (3, 3),
+            "padding": "same"
+        }
+
+        shape = (224, 224, 3)
+
+        x = keras.layers.Input(shape)
+
+        y = keras.layers.Conv2D(64, **options)(x)
+        y = keras.layers.Conv2D(64, **options)(y)
+
+        y = keras.layers.MaxPooling2D(strides=(2, 2))(y)
+
+        y = keras.layers.Conv2D(128, **options)(y)
+        y = keras.layers.Conv2D(128, **options)(y)
+
+        y = keras.layers.MaxPooling2D(strides=(2, 2))(y)
+
+        y = keras.layers.Conv2D(256, **options)(y)
+        y = keras.layers.Conv2D(256, **options)(y)
+        y = keras.layers.Conv2D(256, **options)(y)
+        y = keras.layers.Conv2D(256, **options)(y)
+
+        y = keras.layers.MaxPooling2D(strides=(2, 2))(y)
+
+        y = keras.layers.Conv2D(512, **options)(y)
+        y = keras.layers.Conv2D(512, **options)(y)
+        y = keras.layers.Conv2D(512, **options)(y)
+        y = keras.layers.Conv2D(512, **options)(y)
+
+        y = keras.layers.MaxPooling2D(strides=(2, 2))(y)
+
+        y = keras.layers.Conv2D(512, **options)(y)
+        y = keras.layers.Conv2D(512, **options)(y)
+        y = keras.layers.Conv2D(512, **options)(y)
+
+        y = keras.layers.Conv2D(512, **options)(y)
+
+        a = keras.layers.Conv2D(9 * 4, (1, 1))(y)
+
+        b = keras.layers.Conv2D(9 * 2, (1, 1), activation="sigmoid")(y)
+
+        y = keras_rcnn.layers.object_detection.ObjectProposal(300)([a, b])
+
+        model = keras.models.Model(x, y)
+
+        model.compile("sgd", "mse")
 
-        a = numpy.random.rand(1, 14, 14, 4 * 9)
-        b = numpy.random.rand(1, 14, 14, 2 * 9)
+        image = numpy.random.rand(1, 224, 224, 3)
 
-        prediction = object_proposal_model.predict([a, b])
+        prediction = model.predict(image)
 
-        assert prediction.shape == (1, 100, 4)
+        assert prediction.shape == (1, 300, 4)
 
     def test_compute_output_shape(self, object_proposal_layer):
         assert object_proposal_layer.compute_output_shape((14, 14)) == (None, 300, 4)