Enable option for fixed size crops in get_image_features (#98)

* First pass at code for fixed size crops vs crop and resize

* Fix error in X_padding

* Fixed missing callto X_padded and y_padded

* Fixed X_padded and y_padded

* Add test for fixed crop mode in get_image_features

* Adjusted norm flag so that zeroing out non label information only happens if the norm flag is true

* Update deepcell_tracking/utils.py

* Updating docstring for get_image_features

deepcell_tracking/utils.py

@@ -409,18 +409,27 @@ def is_valid_lineage(y, lineage):
     return is_valid  # if unchanged, all cell lineages are valid!
 
 
-def get_image_features(X, y, appearance_dim=32):
+def get_image_features(X, y, appearance_dim=32, crop_mode='resize', norm=True):
     """Return features for every object in the array.
 
     Args:
         X (np.array): a 3D numpy array of raw data of shape (x, y, c).
         y (np.array): a 3D numpy array of integer labels of shape (x, y, 1).
         appearance_dim (int): The resized shape of the appearance feature.
+        crop_mode (str): Whether to do a fixed crop or to crop and resize
+            to create the appearance features
+        norm (bool): Whether to remove non cell features and normalize the
+            foreground pixels by zero-meaning and dividing by the standard
+            deviation. Applies to fixed crop mode only.
 
     Returns:
         dict: A dictionary of feature names to np.arrays of shape
             (n, c) or (n, x, y, c) where n is the number of objects.
     """
+
+    if crop_mode not in ['resize', 'fixed']:
+        raise ValueError('crop_mode must be either resize or fixed')
+
     appearance_dim = int(appearance_dim)
 
     # each feature will be ordered based on the label.
@@ -432,9 +441,22 @@ def get_image_features(X, y, appearance_dim=32):
     appearances = np.zeros((num_labels, appearance_dim,
                             appearance_dim, X.shape[-1]), dtype='float32')
 
+    if crop_mode == 'fixed':
+        # Zero-pad the X array for fixed crop mode
+        pad_width = ((appearance_dim, appearance_dim),
+                     (appearance_dim, appearance_dim),
+                     (0, 0))
+        X_padded = np.pad(X, pad_width=pad_width)
+        y_padded = np.pad(y, pad_width=pad_width)
+
+        props = regionprops(y_padded[..., 0], cache=False)
+
     # iterate over all objects in y
-    props = regionprops(y[..., 0], cache=False)
+    if crop_mode == 'resize':
+        props = regionprops(y[..., 0], cache=False)
+
     for i, prop in enumerate(props):
+
         # Get label
         labels[i] = prop.label
 
@@ -450,23 +472,43 @@ def get_image_features(X, y, appearance_dim=32):
         ])
         morphologies[i] = morphology
 
-        # Get appearance
-        minr, minc, maxr, maxc = prop.bbox
-        appearance = np.copy(X[minr:maxr, minc:maxc, :])
-        resize_shape = (appearance_dim, appearance_dim)
-        appearance = resize(appearance, resize_shape)
-        appearances[i] = appearance
-
-    # Get adjacency matrix
-    # distance = cdist(centroids, centroids, metric='euclidean') < distance_threshold
-    # adj_matrix = distance.astype('float32')
+        if crop_mode == 'resize':
+            # Get appearance
+            minr, minc, maxr, maxc = prop.bbox
+            appearance = np.copy(X[minr:maxr, minc:maxc, :])
+            resize_shape = (appearance_dim, appearance_dim)
+            appearance = resize(appearance, resize_shape)
+            appearances[i] = appearance
+
+        if crop_mode == 'fixed':
+            cent = np.array(prop.centroid)
+            delta = appearance_dim // 2
+            minr = int(cent[0]) - delta
+            maxr = int(cent[0]) + delta
+            minc = int(cent[1]) - delta
+            maxc = int(cent[1]) + delta
+
+            app = np.copy(X_padded[minr:maxr, minc:maxc, :])
+            label = np.copy(y_padded[minr:maxr, minc:maxc])
+
+            if norm:
+                # Use label as a mask to zero out non-label information
+                app = app * (label == prop.label)
+                idx = np.nonzero(app)
+
+                # Check data and normalize
+                if len(idx) > 0:
+                    mean = np.mean(app[idx])
+                    std = np.std(app[idx])
+                    app[idx] = (app[idx] - mean) / std
+
+            appearances[i] = app
 
     return {
         'appearances': appearances,
         'centroids': centroids,
         'labels': labels,
         'morphologies': morphologies,
-        # 'adj_matrix': adj_matrix,
     }
 
 

deepcell_tracking/utils_test.py

@@ -367,6 +367,13 @@ def test_get_image_features(self):
         assert labels.shape == expected_shape
         np.testing.assert_array_equal(labels, np.array(list(range(1, num_labels + 1))))
 
+        # test appearance - fixed crop
+        features = utils.get_image_features(X, y, appearance_dim,
+                                            crop_mode='fixed', norm=True)
+        appearances = features['appearances']
+        expected_shape = (num_labels, appearance_dim, appearance_dim, X.shape[-1])
+        assert appearances.shape == expected_shape
+
     def test_trks_stats(self):
         # Test bad extension
         with pytest.raises(ValueError):