convert masks to ndarray in transforms (#2030)

mmdet/datasets/pipelines/transforms.py

@@ -156,7 +156,7 @@ class Resize(object):
                     mmcv.imresize(mask, mask_size, interpolation='nearest')
                     for mask in results[key]
                 ]
-            results[key] = masks
+            results[key] = np.stack(masks)
 
     def _resize_seg(self, results):
         for key in results.get('seg_fields', []):
@@ -245,10 +245,10 @@ class RandomFlip(object):
                                               results['flip_direction'])
             # flip masks
             for key in results.get('mask_fields', []):
-                results[key] = [
+                results[key] = np.stack([
                     mmcv.imflip(mask, direction=results['flip_direction'])
                     for mask in results[key]
-                ]
+                ])
 
             # flip segs
             for key in results.get('seg_fields', []):
@@ -410,7 +410,7 @@ class RandomCrop(object):
                     gt_mask = results['gt_masks'][i][crop_y1:crop_y2,
                                                      crop_x1:crop_x2]
                     valid_gt_masks.append(gt_mask)
-                results['gt_masks'] = valid_gt_masks
+                results['gt_masks'] = np.stack(valid_gt_masks)
 
         return results
 
@@ -586,7 +586,7 @@ class Expand(object):
                                       0).astype(mask.dtype)
                 expand_mask[top:top + h, left:left + w] = mask
                 expand_gt_masks.append(expand_mask)
-            results['gt_masks'] = expand_gt_masks
+            results['gt_masks'] = np.stack(expand_gt_masks)
 
         # not tested
         if 'gt_semantic_seg' in results:
@@ -678,10 +678,10 @@ class MinIoURandomCrop(object):
                         results['gt_masks'][i] for i in range(len(mask))
                         if mask[i]
                     ]
-                    results['gt_masks'] = [
+                    results['gt_masks'] = np.stack([
                         gt_mask[patch[1]:patch[3], patch[0]:patch[2]]
                         for gt_mask in valid_masks
-                    ]
+                    ])
 
                 # not tested
                 if 'gt_semantic_seg' in results: