Fix power of mask for inference

inference/model_pipeline.py

@@ -14,24 +14,28 @@ class VSNetModelPipeline:
         self.no_detected_resize = no_detected_resize
         self.model = model
         self.face_detector = face_detector
-        self.mask = self.create_circular_mask(face_detector.target_size, face_detector.target_size, power=1 / 4)
+        self.mask = self.create_circular_mask(face_detector.target_size, face_detector.target_size)
 
     @staticmethod
-    def create_circular_mask(h, w, center=None, power=None):
-
-        if center is None:  # use the middle of the image
-            center = (int(w / 2), int(h / 2))
+    def create_circular_mask(h, w, power=None, clipping_coef=0.85):
+        center = (int(w / 2), int(h / 2))
 
         Y, X = np.ogrid[:h, :w]
         dist_from_center = np.sqrt((X - center[0]) ** 2 + (Y - center[1]) ** 2)
-        dist_from_center = np.clip(dist_from_center, a_min=0, a_max=max(h / 2, w / 2))
-        dist_from_center = 1 - dist_from_center / np.max(dist_from_center)
+        print(dist_from_center.max(), dist_from_center.min())
+        clipping_radius = min((h - center[0]), (w - center[1])) * clipping_coef
+        max_size = max((h - center[0]), (w - center[1]))
+        dist_from_center[dist_from_center < clipping_radius] = clipping_radius
+        dist_from_center[dist_from_center > max_size] = max_size
+        max_distance, min_distance = np.max(dist_from_center), np.min(dist_from_center)
+        dist_from_center = 1 - (dist_from_center - min_distance) / (max_distance - min_distance)
         if power is not None:
             dist_from_center = np.power(dist_from_center, power)
         dist_from_center = np.stack([dist_from_center] * 3, axis=2)
         # mask = dist_from_center <= radius
         return dist_from_center
 
+
     @staticmethod
     def resize_size(image, size=720, always_apply=True):
         h, w, c = np.shape(image)