[Enchance] support infererence with padding (#1607)

* [Enchance] support infererence with padding

* limite pad after flip when inference

* add test code

mmseg/datasets/pipelines/test_time_aug.py

@@ -57,6 +57,14 @@ class MultiScaleFlipAug(object):
                  img_ratios=None,
                  flip=False,
                  flip_direction='horizontal'):
+        if flip:
+            trans_index = {
+                key['type']: index
+                for index, key in enumerate(transforms)
+            }
+            if 'RandomFlip' in trans_index and 'Pad' in trans_index:
+                assert trans_index['RandomFlip'] < trans_index['Pad'], \
+                    'Pad must be executed after RandomFlip when flip is True'
         self.transforms = Compose(transforms)
         if img_ratios is not None:
             img_ratios = img_ratios if isinstance(img_ratios,

mmseg/models/segmentors/encoder_decoder.py

@@ -189,6 +189,9 @@ class EncoderDecoder(BaseSegmentor):
                 count_mat.cpu().detach().numpy()).to(device=img.device)
         preds = preds / count_mat
         if rescale:
+            # remove padding area
+            resize_shape = img_meta[0]['img_shape'][:2]
+            preds = preds[:, :, :resize_shape[0], :resize_shape[1]]
             preds = resize(
                 preds,
                 size=img_meta[0]['ori_shape'][:2],
@@ -206,6 +209,9 @@ class EncoderDecoder(BaseSegmentor):
             if torch.onnx.is_in_onnx_export():
                 size = img.shape[2:]
             else:
+                # remove padding area
+                resize_shape = img_meta[0]['img_shape'][:2]
+                seg_logit = seg_logit[:, :, :resize_shape[0], :resize_shape[1]]
                 size = img_meta[0]['ori_shape'][:2]
             seg_logit = resize(
                 seg_logit,

tests/test_data/test_tta.py

@@ -149,3 +149,41 @@ def test_multi_scale_flip_aug():
     assert tta_results['scale'] == [(256, 256), (256, 256), (512, 512),
                                     (512, 512), (1024, 1024), (1024, 1024)]
     assert tta_results['flip'] == [False, True, False, True, False, True]
+
+    #  test assertion if flip is True and Pad executed before RandomFlip
+    with pytest.raises(AssertionError):
+        tta_transform = dict(
+            type='MultiScaleFlipAug',
+            img_scale=[(256, 256), (512, 512), (1024, 1024)],
+            img_ratios=None,
+            flip=True,
+            transforms=[
+                dict(type='Resize', keep_ratio=False),
+                dict(type='Pad', size_divisor=32),
+                dict(type='RandomFlip'),
+            ])
+        tta_module = build_from_cfg(tta_transform, PIPELINES)
+
+    tta_transform = dict(
+        type='MultiScaleFlipAug',
+        img_scale=[(256, 256), (512, 512), (1024, 1024)],
+        img_ratios=None,
+        flip=True,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Pad', size_divisor=32),
+        ])
+    tta_module = build_from_cfg(tta_transform, PIPELINES)
+    tta_results = tta_module(results.copy())
+    assert tta_results['scale'] == [(256, 256), (256, 256), (512, 512),
+                                    (512, 512), (1024, 1024), (1024, 1024)]
+    assert tta_results['flip'] == [False, True, False, True, False, True]
+    assert tta_results['img_shape'] == [(144, 256, 3), (144, 256, 3),
+                                        (288, 512, 3), (288, 512, 3),
+                                        (576, 1024, 3), (576, 1024, 3)]
+    assert tta_results['pad_shape'] == [(160, 256, 3), (160, 256, 3),
+                                        (288, 512, 3), (288, 512, 3),
+                                        (576, 1024, 3), (576, 1024, 3)]
+    for i in range(len(tta_results['img'])):
+        assert tta_results['img'][i].shape == tta_results['pad_shape'][i]