[Feature] Add OneMinusNormEditDistance for OCR Task

mmeval/metrics/__init__.py

@@ -18,6 +18,7 @@
 from .mse import MeanSquaredError
 from .multi_label import AveragePrecision, MultiLabelMetric
 from .oid_map import OIDMeanAP
+from .one_minus_norm_edit_distance import OneMinusNormEditDistance
 from .pck_accuracy import JhmdbPCKAccuracy, MpiiPCKAccuracy, PCKAccuracy
 from .proposal_recall import ProposalRecall
 from .psnr import PeakSignalNoiseRatio
@@ -36,7 +37,7 @@
     'StructuralSimilarity', 'SignalNoiseRatio', 'MultiLabelMetric',
     'AveragePrecision', 'AVAMeanAP', 'BLEU', 'DOTAMeanAP',
     'SumAbsoluteDifferences', 'GradientError', 'MattingMeanSquaredError',
-    'ConnectivityError', 'ROUGE'
+    'ConnectivityError', 'ROUGE', 'OneMinusNormEditDistance'
 ]
 
 _deprecated_msg = (

mmeval/metrics/one_minus_norm_edit_distance.py

@@ -0,0 +1,82 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import re
+from typing import TYPE_CHECKING, Dict, List, Sequence
+
+from mmeval.core import BaseMetric
+from mmeval.utils import try_import
+
+if TYPE_CHECKING:
+    from rapidfuzz.distance import Levenshtein
+else:
+    distance = try_import('rapidfuzz.distance')
+    if distance is not None:
+        Levenshtein = distance.Levenshtein
+
+
+class OneMinusNormEditDistance(BaseMetric):
+    """One minus NED metric for text recognition task.
+
+    Args:
+        letter_case (str): There are three options to alter the letter cases
+            - unchanged: Do not change prediction texts and labels.
+            - upper: Convert prediction texts and labels into uppercase
+                     characters.
+            - lower: Convert prediction texts and labels into lowercase
+                     characters.
+            Usually, it only works for English characters. Defaults to
+            'unchanged'.
+        valid_symbol (str): Valid characters. Defaults to
+            '[^A-Z^a-z^0-9^\u4e00-\u9fa5]'.
+
+    Example:
+        >>> from mmeval import OneMinusNormEditDistance
+        >>> metric = OneMinusNormEditDistance()
+        >>> metric(['helL', 'HEL'], ['hello', 'HELLO'])
+        {'1-N.E.D': 0.6}
+        >>> metric = OneMinusNormEditDistance(letter_case='upper')
+        >>> metric(['helL', 'HEL'], ['hello', 'HELLO'])
+        {'1-N.E.D': 0.7}
+    """
+
+    def __init__(self,
+                 letter_case: str = 'unchanged',
+                 valid_symbol: str = '[^A-Z^a-z^0-9^\u4e00-\u9fa5]',
+                 **kwargs):
+        super().__init__(**kwargs)
+
+        assert letter_case in ['unchanged', 'upper', 'lower']
+        self.letter_case = letter_case
+        self.valid_symbol = re.compile(valid_symbol)
+
+    def add(self, predictions: Sequence[str], labels: Sequence[str]):  # type: ignore # yapf: disable # noqa: E501
+        """Process one batch of data and predictions.
+
+        Args:
+            predictions (list[str]): The prediction texts.
+            labels (list[str]): The ground truth texts.
+        """
+        for pred, label in zip(predictions, labels):
+            if self.letter_case in ['upper', 'lower']:
+                pred = getattr(pred, self.letter_case)()
+                label = getattr(label, self.letter_case)()
+            label = self.valid_symbol.sub('', label)
+            pred = self.valid_symbol.sub('', pred)
+            norm_ed = Levenshtein.normalized_distance(pred, label)
+            self._results.append(norm_ed)
+
+    def compute_metric(self, results: List[float]) -> Dict:
+        """Compute the metrics from processed results.
+
+        Args:
+            results (list[float]): The processed results of each batch.
+
+        Returns:
+            dict[str, float]: Nested dicts as results.
+              - 1-N.E.D (float): One minus the normalized edit distance.
+        """
+        gt_word_num = len(results)
+        norm_ed_sum = sum(results)
+        normalized_edit_distance = norm_ed_sum / max(1.0, gt_word_num)
+        eval_res = {}
+        eval_res['1-N.E.D'] = 1.0 - normalized_edit_distance
+        return eval_res

requirements/optional.txt

@@ -1,4 +1,5 @@
 opencv-python!=4.5.5.62,!=4.5.5.64
 pycocotools
+rapidfuzz
 scipy
 shapely

setup.cfg

@@ -18,7 +18,7 @@ SPLIT_BEFORE_EXPRESSION_AFTER_OPENING_PAREN = true
 [codespell]
 skip = *.ipynb
 quiet-level = 3
-ignore-words-list = dota, rouge
+ignore-words-list = dota, rouge, ned
 
 [mypy]
 allow_redefinition = True

tests/test_metrics/test_one_minus_norm_edit_distance.py

@@ -0,0 +1,23 @@
+import pytest
+
+from mmeval import OneMinusNormEditDistance
+
+
+def test_init():
+    with pytest.raises(AssertionError):
+        OneMinusNormEditDistance(letter_case='fake')
+
+
+def test_one_minus_norm_edit_distance_metric():
+    metric = OneMinusNormEditDistance(letter_case='lower')
+    res = metric(['helL', 'HEL'], ['hello', 'HELLO'])
+    assert abs(res['1-N.E.D'] - 0.7) < 1e-7
+    metric = OneMinusNormEditDistance(letter_case='upper')
+    res = metric(['helL', 'HEL'], ['hello', 'HELLO'])
+    assert abs(res['1-N.E.D'] - 0.7) < 1e-7
+    metric = OneMinusNormEditDistance()
+    res = metric(['helL', 'HEL'], ['hello', 'HELLO'])
+    assert abs(res['1-N.E.D'] - 0.6) < 1e-7
+
+
+test_one_minus_norm_edit_distance_metric()