Fix mask resizing when mask is all FG and no BG.

If the mask is a rectangle, its mini-mask pixels
become all 1s, and it has no zeros. Scipy imresize
fails on that and returns 0s instead. This update
fixes the problem.

model.py

@@ -20,7 +20,7 @@ import logging
 from collections import OrderedDict
 import multiprocessing
 import numpy as np
-import scipy.misc
+import skimage.transform
 import tensorflow as tf
 import keras
 import keras.backend as K
@@ -1408,16 +1408,16 @@ def build_detection_targets(rpn_rois, gt_class_ids, gt_boxes, gt_masks, config):
             gt_h = gt_y2 - gt_y1
             # Resize mini mask to size of GT box
             placeholder[gt_y1:gt_y2, gt_x1:gt_x2] = \
-                np.round(scipy.misc.imresize(class_mask.astype(float), (gt_h, gt_w),
-                                             interp='nearest') / 255.0).astype(bool)
+                np.round(skimage.transform.resize(
+                    class_mask, (gt_h, gt_w), order=1, mode="reflect")).astype(bool)
             # Place the mini batch in the placeholder
             class_mask = placeholder
 
         # Pick part of the mask and resize it
         y1, x1, y2, x2 = rois[i].astype(np.int32)
         m = class_mask[y1:y2, x1:x2]
-        mask = scipy.misc.imresize(
-            m.astype(float), config.MASK_SHAPE, interp='nearest') / 255.0
+        mask = skimage.transform.resize(m, config.MASK_SHAPE,
+                                        order=1, mode="reflect")
         masks[i, :, :, class_id] = mask
 
     return rois, roi_gt_class_ids, bboxes, masks

train_shapes.ipynb

@@ -276,7 +276,7 @@
     "            occlusion = np.logical_and(occlusion, np.logical_not(mask[:, :, i]))\n",
     "        # Map class names to class IDs.\n",
     "        class_ids = np.array([self.class_names.index(s[0]) for s in shapes])\n",
-    "        return mask, class_ids.astype(np.int32)\n",
+    "        return mask.astype(np.bool), class_ids.astype(np.int32)\n",
     "\n",
     "    def draw_shape(self, image, shape, dims, color):\n",
     "        \"\"\"Draws a shape from the given specs.\"\"\"\n",

utils.py

@@ -16,6 +16,7 @@ import tensorflow as tf
 import scipy.misc
 import skimage.color
 import skimage.io
+import skimage.transform
 import urllib.request
 import shutil
 import warnings
@@ -445,7 +446,6 @@ def resize_mask(mask, scale, padding):
     padding: Padding to add to the mask in the form
             [(top, bottom), (left, right), (0, 0)]
     """
-    h, w = mask.shape[:2]
     # Suppress warning from scipy 0.13.0, the output shape of zoom() is
     # calculated with round() instead of int()
     with warnings.catch_warnings():
@@ -468,8 +468,8 @@ def minimize_mask(bbox, mask, mini_shape):
         m = m[y1:y2, x1:x2]
         if m.size == 0:
             raise Exception("Invalid bounding box with area of zero")
-        m = scipy.misc.imresize(m.astype(float), mini_shape, interp='bilinear')
-        mini_mask[:, :, i] = np.where(m >= 128, 1, 0)
+        m = skimage.transform.resize(m, mini_shape, order=1, mode="reflect")
+        mini_mask[:, :, i] = np.around(m).astype(np.bool)
     return mini_mask
 
 
@@ -485,8 +485,8 @@ def expand_mask(bbox, mini_mask, image_shape):
         y1, x1, y2, x2 = bbox[i][:4]
         h = y2 - y1
         w = x2 - x1
-        m = scipy.misc.imresize(m.astype(float), (h, w), interp='bilinear')
-        mask[y1:y2, x1:x2, i] = np.where(m >= 128, 1, 0)
+        m = skimage.transform.resize(m, (h, w), order=1, mode="reflect")
+        mask[y1:y2, x1:x2, i] = np.around(m).astype(np.bool)
     return mask
 
 
@@ -505,8 +505,8 @@ def unmold_mask(mask, bbox, image_shape):
     """
     threshold = 0.5
     y1, x1, y2, x2 = bbox
-    mask = scipy.misc.imresize(
-        mask, (y2 - y1, x2 - x1), interp='bilinear').astype(np.float32) / 255.0
+    mask = skimage.transform.resize(mask, (y2 - y1, x2 - x1),
+                                    order=1, mode="reflect")
     mask = np.where(mask >= threshold, 1, 0).astype(np.uint8)
 
     # Put the mask in the right location.
@@ -618,7 +618,7 @@ def compute_ap(gt_boxes, gt_class_ids, gt_masks,
     # Compute IoU overlaps [pred_masks, gt_masks]
     overlaps = compute_overlaps_masks(pred_masks, gt_masks)
 
-    # Loop through ground truth boxes and find matching predictions
+    # Loop through predictions and find matching ground truth boxes
     match_count = 0
     pred_match = np.zeros([pred_boxes.shape[0]])
     gt_match = np.zeros([gt_boxes.shape[0]])