[Fix] Pad boader fix (#1757)

* [Fix] expand pad_val'dim to be same with image

* fix comment

mmcv/transforms/processing.py

@@ -287,7 +287,7 @@ class Pad(BaseTransform):
 
     Args:
         size (tuple, optional): Fixed padding size.
-            Expected padding shape (h, w)Defaults to None.
+            Expected padding shape (w, h). Defaults to None.
         size_divisor (int, optional): The divisor of padded size. Defaults to
             None.
         pad_to_square (bool): Whether to pad the image into a square.
@@ -354,6 +354,8 @@ class Pad(BaseTransform):
             size = (pad_h, pad_w)
         elif self.size is not None:
             size = self.size[::-1]
+        if isinstance(pad_val, int) and results['img'].ndim == 3:
+            pad_val = tuple([pad_val for _ in range(results['img'].shape[2])])
         padded_img = mmcv.impad(
             results['img'],
             shape=size,
@@ -372,7 +374,9 @@ class Pad(BaseTransform):
         ``results['pad_shape']``."""
         if results.get('gt_semantic_seg', None) is not None:
             pad_val = self.pad_val.get('seg', 255)
-
+        if isinstance(pad_val, int) and results['gt_semantic_seg'].ndim == 3:
+            pad_val = tuple(
+                [pad_val for _ in range(results['gt_semantic_seg'].shape[2])])
             results['gt_semantic_seg'] = mmcv.impad(
                 results['gt_semantic_seg'],
                 shape=results['pad_shape'][:2],

tests/test_transforms/test_transforms_processing.py

@@ -171,12 +171,50 @@ class TestPad:
         results = trans(copy.deepcopy(data_info))
         assert (results['img'] == np.ones((1333, 1333, 3))).all()
 
-        # test pad_val
-        new_img = np.zeros((1333, 800, 3))
+        # test pad_val is dict
+        # test rgb image, size=(2000, 2000)
+        trans = Pad(
+            size=(2000, 2000),
+            pad_val=dict(img=(12, 12, 12), seg=(10, 10, 10)))
+        results = trans(copy.deepcopy(data_info))
+        assert (results['img'][1333:2000, 800:2000, :] == 12).all()
+        assert (results['gt_semantic_seg'][1333:2000, 800:2000, :] == 10).all()
+
+        trans = Pad(size=(2000, 2000), pad_val=dict(img=(12, 12, 12)))
+        results = trans(copy.deepcopy(data_info))
+        assert (results['img'][1333:2000, 800:2000, :] == 12).all()
+        assert (results['gt_semantic_seg'][1333:2000,
+                                           800:2000, :] == 255).all()
+
+        # test rgb image, pad_to_square=True
+        trans = Pad(
+            pad_to_square=True,
+            pad_val=dict(img=(12, 12, 12), seg=(10, 10, 10)))
+        results = trans(copy.deepcopy(data_info))
+        assert (results['img'][:, 800:1333, :] == 12).all()
+        assert (results['gt_semantic_seg'][:, 800:1333, :] == 10).all()
+
+        trans = Pad(pad_to_square=True, pad_val=dict(img=(12, 12, 12)))
+        results = trans(copy.deepcopy(data_info))
+        assert (results['img'][:, 800:1333, :] == 12).all()
+        assert (results['gt_semantic_seg'][:, 800:1333, :] == 255).all()
+
+        # test pad_val is int
+        # test rgb image
+        trans = Pad(size=(2000, 2000), pad_val=12)
+        results = trans(copy.deepcopy(data_info))
+        assert (results['img'][1333:2000, 800:2000, :] == 12).all()
+        assert (results['gt_semantic_seg'][1333:2000,
+                                           800:2000, :] == 255).all()
+        # test gray image
+        new_img = np.random.random((1333, 800))
         data_info['img'] = new_img
-        trans = Pad(pad_to_square=True, pad_val=0)
+        new_semantic_seg = np.random.random((1333, 800))
+        data_info['gt_semantic_seg'] = new_semantic_seg
+        trans = Pad(size=(2000, 2000), pad_val=12)
         results = trans(copy.deepcopy(data_info))
-        assert (results['img'] == np.zeros((1333, 1333, 3))).all()
+        assert (results['img'][1333:2000, 800:2000] == 12).all()
+        assert (results['gt_semantic_seg'][1333:2000, 800:2000] == 255).all()
 
     def test_repr(self):
         trans = Pad(pad_to_square=True, size_divisor=11, padding_mode='edge')