update separation lines computation

multi_plankton_separation/api.py

@@ -75,34 +75,6 @@ def get_metadata():
     return meta
 
 
-# def get_train_args():
-#     arg_dict = {
-#         "epoch_num": fields.Int(
-#             required=False,
-#             missing=10,
-#             description="Total number of training epochs",
-#         ),
-#     }
-#     return arg_dict
-
-
-# def train(**kwargs):
-#     """
-#     Dummy training. We just sleep for some number of epochs (1 epoch = 1 second)
-#     mimicking some computation taking place.
-#     We can log some random losses in Tensorboard to mimic monitoring.
-#     """
-#     logdir = BASE_DIR / "runs" / time.strftime("%Y-%m-%d_%H-%M-%S")
-#     writer = SummaryWriter(logdir=logdir)
-#     launch_tensorboard(logdir=logdir)
-#     for epoch in range(kwargs["epoch_num"]):
-#         time.sleep(1.)
-#         writer.add_scalar("scalars/loss", - math.log(epoch + 1), epoch)
-#     writer.close()
-
-#     return {"status": "done", "final accuracy": 0.9}
-
-
 def get_predict_args():
     """
     Get the list of arguments for the predict function
@@ -125,11 +97,16 @@ def get_predict_args():
             enum=list_models,
             description="The model used to perform instance segmentation"
         ),
-        "threshold": fields.Float(
+        "min_mask_score": fields.Float(
             required=False,
             missing=0.9,
             description="The minimum confidence score for a mask to be selected"
         ),
+        "min_mask_value": fields.Float(
+            required=False,
+            missing=0.5,
+            description="The minimum value for a pixel to belong to a mask"
+        ),
         "accept" : fields.Str(
             required=False,
             missing='image/png',
@@ -162,23 +139,31 @@ def predict(**kwargs):
     img = transform(orig_img)
 
     # Get predicted masks
-    pred_masks, pred_masks_probs = get_predicted_masks(model, img, kwargs["threshold"])
+    pred_masks, pred_masks_probs = get_predicted_masks(
+        model, img, kwargs["min_mask_score"]
+    )
 
     # Get sum of masks probabilities and mask centers
     mask_sum = np.zeros(pred_masks[0].shape)
     mask_centers_x = []
     mask_centers_y = []
 
+    # Get sum of masks and mask centers for the watershed
     for mask in pred_masks_probs:
-        mask_sum += mask
+        to_add = mask
+        to_add[to_add < kwargs["min_mask_value"]] = 0
+        mask_sum += to_add
         center_x, center_y = np.unravel_index(np.argmax(mask), mask.shape)
         mask_centers_x.append(center_x)
         mask_centers_y.append(center_y)
 
     mask_centers = zip(mask_centers_x, mask_centers_y)
 
+    # Get silhouette of objects to use as a mask for the watershed
+    binary_img = (img[0, :, :] + img[1, :, :] + img[2, :, :] != 3).numpy().astype(float)
+
     # Apply watershed algorithm
-    watershed_labels = get_watershed_result(mask_sum, mask_centers)
+    watershed_labels = get_watershed_result(mask_sum, mask_centers, mask=binary_img)
 
     # Save output separations
     separation_mask = np.ones(watershed_labels.shape)

multi_plankton_separation/utils.py

@@ -6,7 +6,7 @@
 from torchvision.models.detection.faster_rcnn import FastRCNNPredictor
 from torchvision.models.detection.mask_rcnn import MaskRCNNPredictor
 from scipy import ndimage as ndi
-from skimage.segmentation import watershed
+from skimage.segmentation import watershed, find_boundaries
 
 import multi_plankton_separation.config as cfg
 
@@ -75,26 +75,45 @@ def get_predicted_masks(model, image, score_threshold=0.9, mask_threshold=0.7):
     return pred_masks, pred_masks_probs
 
 
-def get_watershed_result(mask_map, mask_centers):
+def get_watershed_result(mask_map, mask_centers, mask=None):
     """
     Apply the watershed algorithm on the predicted mask map,
     using the mask centers as markers
     """
+    # Prepare watershed markers
     markers_mask = np.zeros(mask_map.shape, dtype=bool)
     for (x, y) in mask_centers:
         markers_mask[x, y] = True
     markers, _ = ndi.label(markers_mask)
 
-    watershed_mask = np.zeros(mask_map.shape, dtype='int64')
-    watershed_mask[mask_map > .01] = 1
+    # Prepare watershed mask
+    if mask is None:
+        watershed_mask = np.zeros(mask_map.shape, dtype='int64')
+        watershed_mask[mask_map > .01] = 1
+    else:
+        watershed_mask = mask
 
+    # Apply watershed
     labels = watershed(
         -mask_map, markers, mask=watershed_mask, watershed_line=False
     )
-    labels_with_lines = watershed(
-        -mask_map, markers, mask=watershed_mask, watershed_line=True
-    )
-    labels_with_lines[labels == 0] = -1
+
+    # Derive separation lines
+    lines = np.zeros(labels.shape)
+    unique_labels = list(np.unique(labels))
+    unique_labels.remove(0)
+
+    for value in unique_labels:
+        single_shape = (labels == value).astype(int)
+        boundaries = find_boundaries(
+            single_shape, connectivity=2, mode='outer', background=0
+        )
+        boundaries[(labels == 0) | (labels == value)] = 0
+        lines[boundaries == 1] = 1
+
+    labels_with_lines = labels
+    labels_with_lines[labels_with_lines == 0] = -1
+    labels_with_lines[lines == 1] = 0
 
     return labels_with_lines