[Fix] flip_ratio in RandomAffine pipeline is inverted (#799)

* fix flip_ratio bug

original: `flip_ratio=1.` means always no flip and `flip_ratio=0.` means always flip
after:      `flip_ratio=1.` means always flip and `flip_ratio=0.` means always no flip

* add some notes to modification

* modify center, add unittests

* clean test codes

* clean test codes

Co-authored-by: wangruohui <12756472+wangruohui@users.noreply.github.com>

mmedit/datasets/pipelines/augmentation.py

@@ -447,7 +447,11 @@ def _get_params(degrees, translate, scale_ranges, shears, flip_ratio,
         else:
             shear = 0.0
 
-        flip = (np.random.rand(2) < flip_ratio).astype(np.int32) * 2 - 1
+        # Because `flip` is used as a multiplier in line 479 and 480,
+        # so -1 stands for flip and 1 stands for no flip. Thus `flip`
+        # should be an 'inverse' flag as the result of the comparison.
+        # See https://github.com/open-mmlab/mmediting/pull/799 for more detail
+        flip = (np.random.rand(2) > flip_ratio).astype(np.int32) * 2 - 1
 
         return angle, translations, scale, shear, flip
 
@@ -521,7 +525,7 @@ def __call__(self, results):
             params = self._get_params(self.degrees, self.translate, self.scale,
                                       self.shear, self.flip_ratio, (h, w))
 
-        center = (w * 0.5 + 0.5, h * 0.5 + 0.5)
+        center = (w * 0.5 - 0.5, h * 0.5 - 0.5)
         M = self._get_inverse_affine_matrix(center, *params)
         M = np.array(M).reshape((2, 3))
 

tests/test_data/test_pipelines/test_augmentation.py

@@ -265,7 +265,28 @@ def test_random_affine(self):
 
         target_keys = ['fg', 'alpha']
 
+        # Test identical transformation
+        alpha = np.random.rand(4, 4).astype(np.float32)
+        fg = np.random.rand(4, 4).astype(np.float32)
+        results = dict(alpha=alpha, fg=fg)
+        random_affine = RandomAffine(['fg', 'alpha'],
+                                     degrees=0, flip_ratio=0.0)
+        random_affine_results = random_affine(results)
+        assert np.allclose(alpha, random_affine_results['alpha'])
+        assert np.allclose(fg, random_affine_results['fg'])
+
+        # Test flip in both direction
+        alpha = np.random.rand(4, 4).astype(np.float32)
+        fg = np.random.rand(4, 4).astype(np.float32)
+        results = dict(alpha=alpha, fg=fg)
+        random_affine = RandomAffine(['fg', 'alpha'],
+                                     degrees=0, flip_ratio=1.0)
+        random_affine_results = random_affine(results)
+        assert np.allclose(alpha[::-1, ::-1], random_affine_results['alpha'])
+        assert np.allclose(fg[::-1, ::-1], random_affine_results['fg'])
+
         # test random affine with different valid setting combinations
+        # only shape are tested
         alpha = np.random.rand(240, 320).astype(np.float32)
         fg = np.random.rand(240, 320).astype(np.float32)
         results = dict(alpha=alpha, fg=fg)