Merge pull request #334 from uraid/master

Fix face and object recognition of rotated images

Author: damianmoore

photonix/classifiers/face/model.py

@@ -17,6 +17,7 @@
 from photonix.classifiers.face.deepface.commons.distance import findEuclideanDistance
 from photonix.classifiers.face.deepface.DeepFace import build_model
 from photonix.photos.utils.redis import redis_connection
+from photonix.photos.utils.metadata import PhotoMetadata
 
 
 GRAPH_FILE = os.path.join('face', 'mtcnn_weights.npy')
@@ -72,8 +73,17 @@ def load_graph(self, graph_file):
     def predict(self, image_file, min_score=0.99):
         # Detects face bounding boxes
         image = Image.open(image_file)
+
         if image.mode != 'RGB':
             image = image.convert('RGB')
+
+        # Perform rotations if decalared in metadata
+        metadata = PhotoMetadata(image_file)
+        if metadata.get('Orientation') in ['Rotate 90 CW', 'Rotate 270 CCW']:
+            image = image.rotate(-90, expand=True)
+        elif metadata.get('Orientation') in ['Rotate 90 CCW', 'Rotate 270 CW']:
+            image = image.rotate(90, expand=True)
+
         image = np.asarray(image)
         results = self.graph['mtcnn'].detect_faces(image)
         return list(filter(lambda f: f['confidence'] > min_score, results))

photonix/classifiers/object/model.py

@@ -11,6 +11,7 @@
 from photonix.classifiers.object.utils import label_map_util
 from photonix.classifiers.base_model import BaseModel
 from photonix.photos.utils.redis import redis_connection
+from photonix.photos.utils.metadata import PhotoMetadata
 
 
 GRAPH_FILE = os.path.join('object', 'ssd_mobilenet_v2_oid_v4_2018_12_12_frozen_inference_graph.pb')
@@ -115,8 +116,17 @@ def format_output(self, output_dict, min_score):
 
     def predict(self, image_file, min_score=0.1):
         image = Image.open(image_file)
+
         if image.mode != 'RGB':
             image = image.convert('RGB')
+
+        # Perform rotations if decalared in metadata
+        metadata = PhotoMetadata(image_file)
+        if metadata.get('Orientation') in ['Rotate 90 CW', 'Rotate 270 CCW']:
+            image = image.rotate(-90, expand=True)
+        elif metadata.get('Orientation') in ['Rotate 90 CCW', 'Rotate 270 CW']:
+            image = image.rotate(90, expand=True)
+
         # the array based representation of the image will be used later in order to prepare the
         # result image with boxes and labels on it.
         image_np = self.load_image_into_numpy_array(image)