[Fix] Fix reduce_zero_label in evaluation (#2504)

## Motivation

Through this PR, I (1) fix a bug, and (2) perform some associated cleanup, and (3) add a unit test. The bug occurs during evaluation when two options -- `reduce_zero_label=True`, and custom classes are used. The bug was that the `reduce_zero_label` is not properly propagated (see details below).

## Modification

1. **Bugfix**

The bug occurs [in the initialization of `CustomDataset`](5d49918b3c/mmseg/datasets/custom.py (L108-L110)) where the `reduce_zero_label` flag is not propagated to its member `self.gt_seg_map_loader_cfg`:

```python
self.gt_seg_map_loader = LoadAnnotations(
) if gt_seg_map_loader_cfg is None else LoadAnnotations(
    **gt_seg_map_loader_cfg)
```

Because the `reduce_zero_label` flag was not being propagated, the zero label reduction was being [unnecessarily and explicitly duplicated during the evaluation](5d49918b3c/mmseg/core/evaluation/metrics.py (L66-L69)).

As pointed in a previous PR (#2500), `reduce_zero_label` must occur before applying the `label_map`. Due to this bug, the order gets reversed when both features are used simultaneously.

This has been fixed to:

```python
self.gt_seg_map_loader = LoadAnnotations(
    reduce_zero_label=reduce_zero_label, **gt_seg_map_loader_cfg)
```

2. **Cleanup**

Due to the bug fix, since both `reduce_zero_label` and `label_map` are being applied in `get_gt_seg_map_by_idx()` (i.e. `LoadAnnotations.__call__()`), the evaluation does not need perform them anymore. However, for backwards compatibility, the evaluation keeps previous input arguments.

This was pointed out for `label_map` in a previous issue (#1415) that the `label_map` should  not be applied in the evaluation. This was handled by [passing an empty dict](5d49918b3c/mmseg/datasets/custom.py (L306-L311)):

```python
# as the labels has been converted when dataset initialized
# in `get_palette_for_custom_classes ` this `label_map`
# should be `dict()`, see
# https://github.com/open-mmlab/mmsegmentation/issues/1415
# for more ditails
label_map=dict(),
reduce_zero_label=self.reduce_zero_label))
```

Similar to this, I now also set `reduce_label=False` since it is now also being handled by `get_gt_seg_map_by_idx()` (i.e. `LoadAnnotations.__call__()`).

3. **Unit test**

I've added a unit test that tests the `CustomDataset.pre_eval()` function when `reduce_zero_label=True` and custom classes are used. The test fails on the original `master` branch but passes with this fix.

## BC-breaking (Optional)

I do not anticipate this change braking any backward-compatibility.

## Checklist

- [x] Pre-commit or other linting tools are used to fix the potential lint issues.
  - _I've fixed all linting/pre-commit errors._
- [x] The modification is covered by complete unit tests. If not, please add more unit test to ensure the correctness.
  - _I've added a test that passes when the fix is introduced, and fails on the original master branch._
- [x] If the modification has potential influence on downstream projects, this PR should be tested with downstream projects, like MMDet or MMDet3D.
  - _I don't think this change affects MMDet or MMDet3D._
- [x] The documentation has been modified accordingly, like docstring or example tutorials.
  - _This change fixes an existing bug and doesn't require modifying any documentation/docstring._

mmseg/datasets/custom.py

@@ -66,8 +66,8 @@ class CustomDataset(Dataset):
             The palette of segmentation map. If None is given, and
             self.PALETTE is None, random palette will be generated.
             Default: None
-        gt_seg_map_loader_cfg (dict, optional): build LoadAnnotations to
+        gt_seg_map_loader_cfg (dict): build LoadAnnotations to load gt for
-            load gt for evaluation, load from disk by default. Default: None.
+            evaluation, load from disk by default. Default: ``dict()``.
         file_client_args (dict): Arguments to instantiate a FileClient.
             See :class:`mmcv.fileio.FileClient` for details.
             Defaults to ``dict(backend='disk')``.
@@ -90,7 +90,7 @@ class CustomDataset(Dataset):
                  reduce_zero_label=False,
                  classes=None,
                  palette=None,
-                 gt_seg_map_loader_cfg=None,
+                 gt_seg_map_loader_cfg=dict(),
                  file_client_args=dict(backend='disk')):
         self.pipeline = Compose(pipeline)
         self.img_dir = img_dir
@@ -106,8 +106,7 @@ class CustomDataset(Dataset):
         self.CLASSES, self.PALETTE = self.get_classes_and_palette(
             classes, palette)
         self.gt_seg_map_loader = LoadAnnotations(
-        ) if gt_seg_map_loader_cfg is None else LoadAnnotations(
+            reduce_zero_label=reduce_zero_label, **gt_seg_map_loader_cfg)
-            **gt_seg_map_loader_cfg)
 
         self.file_client_args = file_client_args
         self.file_client = mmcv.FileClient.infer_client(self.file_client_args)
@@ -303,13 +302,16 @@ class CustomDataset(Dataset):
                     seg_map,
                     len(self.CLASSES),
                     self.ignore_index,
-                    # as the labels has been converted when dataset initialized
+                    # as the label map has already been applied and zero label
-                    # in `get_palette_for_custom_classes ` this `label_map`
+                    # has already been reduced by get_gt_seg_map_by_idx() i.e.
-                    # should be `dict()`, see
+                    # LoadAnnotations.__call__(), these operations should not
+                    # be duplicated. See the following issues/PRs:
                     # https://github.com/open-mmlab/mmsegmentation/issues/1415
-                    # for more ditails
+                    # https://github.com/open-mmlab/mmsegmentation/pull/1417
+                    # https://github.com/open-mmlab/mmsegmentation/pull/2504
+                    # for more details
                     label_map=dict(),
-                    reduce_zero_label=self.reduce_zero_label))
+                    reduce_zero_label=False))
 
         return pre_eval_results
 
@@ -427,7 +429,7 @@ class CustomDataset(Dataset):
                 self.ignore_index,
                 metric,
                 label_map=dict(),
-                reduce_zero_label=self.reduce_zero_label)
+                reduce_zero_label=False)
         # test a list of pre_eval_results
         else:
             ret_metrics = pre_eval_to_metrics(results, metric)

tests/test_data/test_dataset.py

@@ -365,6 +365,69 @@ def test_custom_dataset():
     assert not np.isnan(eval_results['mRecall'])
 
 
+def test_custom_dataset_pre_eval():
+    """Test pre-eval function of custom dataset with reduce zero label and
+    removed classes.
+
+    The GT segmentation contain 4 classes: "A", "B", "C", "D", as well as
+    a zero label. Therefore, the labels go from 0 to 4.
+
+    Then, we will remove class "C" while instantiating the dataset. Therefore,
+    pre-eval must reduce the zero label and also apply label_map in the correct
+    order.
+    """
+
+    # create a dummy dataset on disk
+    img = np.random.rand(10, 10)
+    ann = np.zeros_like(img)
+    ann[2:4, 2:4] = 1
+    ann[2:4, 6:8] = 2
+    ann[6:8, 2:4] = 3
+    ann[6:8, 6:8] = 4
+
+    tmp_dir = tempfile.TemporaryDirectory()
+    img_path = osp.join(tmp_dir.name, 'img', '00000.jpg')
+    ann_path = osp.join(tmp_dir.name, 'ann', '00000.png')
+
+    import mmcv
+    mmcv.imwrite(img, img_path)
+    mmcv.imwrite(ann, ann_path)
+
+    class FourClassDatasetWithZeroLabel(CustomDataset):
+        CLASSES = ['A', 'B', 'C', 'D']  # 4 classes
+        PALETTE = [(0, 0, 0)] * 4  # dummy palette
+
+    # with img_dir, ann_dir, split
+    dataset = FourClassDatasetWithZeroLabel(
+        [],
+        classes=['A', 'B', 'D'],  # original classes with class "C" removed
+        reduce_zero_label=True,  # reduce zero label set to True
+        data_root=osp.join(osp.dirname(__file__), tmp_dir.name),
+        img_dir='img/',
+        ann_dir='ann/',
+        img_suffix='.jpg',
+        seg_map_suffix='.png')
+    assert len(dataset) == 1
+
+    # there are three classes ("A", "B", "D") that the network predicts
+    perfect_pred = np.zeros([10, 10], dtype=np.int64)
+    perfect_pred[2:4, 2:4] = 0  # 'A': 1 reduced to 0 that maps to 0
+    perfect_pred[2:4, 6:8] = 1  # 'B': 2 reduced to 1 that maps to 1
+    perfect_pred[6:8, 2:4] = 0  # 'C': 3 reduced to 2 that maps to -1, ignored
+    perfect_pred[6:8, 6:8] = 2  # 'D': 4 reduced to 3 that maps to 2
+
+    results = dataset.pre_eval([perfect_pred], [0])
+    from mmseg.core.evaluation.metrics import pre_eval_to_metrics
+    eval_results = pre_eval_to_metrics(results, ['mIoU', 'mDice', 'mFscore'])
+
+    # the results should be perfect
+    for metric in 'IoU', 'aAcc', 'Acc', 'Dice', 'Fscore', 'Precision', \
+                  'Recall':
+        assert (eval_results[metric] == 1.0).all()
+
+    tmp_dir.cleanup()
+
+
 @pytest.mark.parametrize('separate_eval', [True, False])
 def test_eval_concat_custom_dataset(separate_eval):
     img_norm_cfg = dict(