File size: 12,682 Bytes
9bf4bd7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
# Copyright (c) OpenMMLab. All rights reserved.
import warnings
from typing import Any, Dict, List, Optional, Tuple, Union

import imgaug
import imgaug.augmenters as iaa
import numpy as np
import torchvision.transforms as torchvision_transforms
from mmcv.transforms import Compose
from mmcv.transforms.base import BaseTransform
from PIL import Image

from mmocr.registry import TRANSFORMS
from mmocr.utils import poly2bbox


@TRANSFORMS.register_module()
class ImgAugWrapper(BaseTransform):
    """A wrapper around imgaug https://github.com/aleju/imgaug.

    Find available augmenters at
    https://imgaug.readthedocs.io/en/latest/source/overview_of_augmenters.html.

    Required Keys:

    - img
    - gt_polygons (optional for text recognition)
    - gt_bboxes (optional for text recognition)
    - gt_bboxes_labels (optional for text recognition)
    - gt_ignored (optional for text recognition)
    - gt_texts (optional)

    Modified Keys:

    - img
    - gt_polygons (optional for text recognition)
    - gt_bboxes (optional for text recognition)
    - gt_bboxes_labels (optional for text recognition)
    - gt_ignored (optional for text recognition)
    - img_shape (optional)
    - gt_texts (optional)

    Args:
        args (list[list or dict]], optional): The argumentation list. For
            details, please refer to imgaug document. Take
            args=[['Fliplr', 0.5], dict(cls='Affine', rotate=[-10, 10]),
            ['Resize', [0.5, 3.0]]] as an example. The args horizontally flip
            images with probability 0.5, followed by random rotation with
            angles in range [-10, 10], and resize with an independent scale in
            range [0.5, 3.0] for each side of images. Defaults to None.
        fix_poly_trans (dict): The transform configuration to fix invalid
            polygons. Set it to None if no fixing is needed.
            Defaults to dict(type='FixInvalidPolygon').
    """

    def __init__(
        self,
        args: Optional[List[Union[List, Dict]]] = None,
        fix_poly_trans: Optional[dict] = dict(type='FixInvalidPolygon')
    ) -> None:
        assert args is None or isinstance(args, list) and len(args) > 0
        if args is not None:
            for arg in args:
                assert isinstance(arg, (list, dict)), \
                    'args should be a list of list or dict'
        self.args = args
        self.augmenter = self._build_augmentation(args)
        self.fix_poly_trans = fix_poly_trans
        if fix_poly_trans is not None:
            self.fix = TRANSFORMS.build(fix_poly_trans)

    def transform(self, results: Dict) -> Dict:
        """Transform the image and annotation data.

        Args:
            results (dict): Result dict containing the data to transform.

        Returns:
            dict: The transformed data.
        """
        # img is bgr
        image = results['img']
        aug = None
        ori_shape = image.shape

        if self.augmenter:
            aug = self.augmenter.to_deterministic()
            if not self._augment_annotations(aug, ori_shape, results):
                return None
            results['img'] = aug.augment_image(image)
            results['img_shape'] = (results['img'].shape[0],
                                    results['img'].shape[1])
        if getattr(self, 'fix', None) is not None:
            results = self.fix(results)
        return results

    def _augment_annotations(self, aug: imgaug.augmenters.meta.Augmenter,
                             ori_shape: Tuple[int,
                                              int], results: Dict) -> Dict:
        """Augment annotations following the pre-defined augmentation sequence.

        Args:
            aug (imgaug.augmenters.meta.Augmenter): The imgaug augmenter.
            ori_shape (tuple[int, int]): The ori_shape of the original image.
            results (dict): Result dict containing annotations to transform.

        Returns:
            bool: Whether the transformation has been successfully applied. If
            the transform results in empty polygon/bbox annotations, return
            False.
        """
        # Assume co-existence of `gt_polygons`, `gt_bboxes` and `gt_ignored`
        # for text detection
        if 'gt_polygons' in results:

            # augment polygons
            transformed_polygons, removed_poly_inds = self._augment_polygons(
                aug, ori_shape, results['gt_polygons'])
            if len(transformed_polygons) == 0:
                return False
            results['gt_polygons'] = transformed_polygons

            # remove instances that are no longer inside the augmented image
            results['gt_bboxes_labels'] = np.delete(
                results['gt_bboxes_labels'], removed_poly_inds, axis=0)
            results['gt_ignored'] = np.delete(
                results['gt_ignored'], removed_poly_inds, axis=0)
            # TODO: deal with gt_texts corresponding to clipped polygons
            if 'gt_texts' in results:
                results['gt_texts'] = [
                    text for i, text in enumerate(results['gt_texts'])
                    if i not in removed_poly_inds
                ]

            # Generate new bboxes
            bboxes = [poly2bbox(poly) for poly in transformed_polygons]
            results['gt_bboxes'] = np.zeros((0, 4), dtype=np.float32)
            if len(bboxes) > 0:
                results['gt_bboxes'] = np.stack(bboxes)

        return True

    def _augment_polygons(self, aug: imgaug.augmenters.meta.Augmenter,
                          ori_shape: Tuple[int, int], polys: List[np.ndarray]
                          ) -> Tuple[List[np.ndarray], List[int]]:
        """Augment polygons.

        Args:
            aug (imgaug.augmenters.meta.Augmenter): The imgaug augmenter.
            ori_shape (tuple[int, int]): The shape of the original image.
            polys (list[np.ndarray]): The polygons to be augmented.

        Returns:
            tuple(list[np.ndarray], list[int]): The augmented polygons, and the
            indices of polygons removed as they are out of the augmented image.
        """
        imgaug_polys = []
        for poly in polys:
            poly = poly.reshape(-1, 2)
            imgaug_polys.append(imgaug.Polygon(poly))
        imgaug_polys = aug.augment_polygons(
            [imgaug.PolygonsOnImage(imgaug_polys, shape=ori_shape)])[0]

        new_polys = []
        removed_poly_inds = []
        for i, poly in enumerate(imgaug_polys.polygons):
            # Sometimes imgaug may produce some invalid polygons with no points
            if not poly.is_valid or poly.is_out_of_image(imgaug_polys.shape):
                removed_poly_inds.append(i)
                continue
            new_poly = []
            try:
                poly = poly.clip_out_of_image(imgaug_polys.shape)[0]
            except Exception as e:
                warnings.warn(f'Failed to clip polygon out of image: {e}')
            for point in poly:
                new_poly.append(np.array(point, dtype=np.float32))
            new_poly = np.array(new_poly, dtype=np.float32).flatten()
            # Under some conditions, imgaug can generate "polygon" with only
            # two points, which is not a valid polygon.
            if len(new_poly) <= 4:
                removed_poly_inds.append(i)
                continue
            new_polys.append(new_poly)

        return new_polys, removed_poly_inds

    def _build_augmentation(self, args, root=True):
        """Build ImgAugWrapper augmentations.

        Args:
            args (dict): Arguments to be passed to imgaug.
            root (bool): Whether it's building the root augmenter.

        Returns:
            imgaug.augmenters.meta.Augmenter: The built augmenter.
        """
        if args is None:
            return None
        if isinstance(args, (int, float, str)):
            return args
        if isinstance(args, list):
            if root:
                sequence = [
                    self._build_augmentation(value, root=False)
                    for value in args
                ]
                return iaa.Sequential(sequence)
            arg_list = [self._to_tuple_if_list(a) for a in args[1:]]
            return getattr(iaa, args[0])(*arg_list)
        if isinstance(args, dict):
            if 'cls' in args:
                cls = getattr(iaa, args['cls'])
                return cls(
                    **{
                        k: self._to_tuple_if_list(v)
                        for k, v in args.items() if not k == 'cls'
                    })
            else:
                return {
                    key: self._build_augmentation(value, root=False)
                    for key, value in args.items()
                }
        raise RuntimeError('unknown augmenter arg: ' + str(args))

    def _to_tuple_if_list(self, obj: Any) -> Any:
        """Convert an object into a tuple if it is a list."""
        if isinstance(obj, list):
            return tuple(obj)
        return obj

    def __repr__(self):
        repr_str = self.__class__.__name__
        repr_str += f'(args = {self.args}, '
        repr_str += f'fix_poly_trans = {self.fix_poly_trans})'
        return repr_str


@TRANSFORMS.register_module()
class TorchVisionWrapper(BaseTransform):
    """A wrapper around torchvision transforms. It applies specific transform
    to ``img`` and updates ``height`` and ``width`` accordingly.

    Required Keys:

    - img (ndarray): The input image.

    Modified Keys:

    - img (ndarray): The modified image.
    - img_shape (tuple(int, int)): The shape of the image in (height, width).


    Warning:
        This transform only affects the image but not its associated
        annotations, such as word bounding boxes and polygons. Therefore,
        it may only be applicable to text recognition tasks.

    Args:
        op (str): The name of any transform class in
            :func:`torchvision.transforms`.
        **kwargs: Arguments that will be passed to initializer of torchvision
            transform.
    """

    def __init__(self, op: str, **kwargs) -> None:
        assert isinstance(op, str)
        obj_cls = getattr(torchvision_transforms, op)
        self.torchvision = obj_cls(**kwargs)
        self.op = op
        self.kwargs = kwargs

    def transform(self, results):
        """Transform the image.

        Args:
            results (dict): Result dict from the data loader.

        Returns:
            dict: Transformed results.
        """
        assert 'img' in results
        # BGR -> RGB
        img = results['img'][..., ::-1]
        img = Image.fromarray(img)
        img = self.torchvision(img)
        img = np.asarray(img)
        img = img[..., ::-1]
        results['img'] = img
        results['img_shape'] = img.shape[:2]
        return results

    def __repr__(self):
        repr_str = self.__class__.__name__
        repr_str += f'(op = {self.op}'
        for k, v in self.kwargs.items():
            repr_str += f', {k} = {v}'
        repr_str += ')'
        return repr_str


@TRANSFORMS.register_module()
class ConditionApply(BaseTransform):
    """Apply transforms according to the condition. If the condition is met,
    true_transforms will be applied, otherwise false_transforms will be
    applied.

    Args:
        condition (str): The string that can be evaluated to a boolean value.
        true_transforms (list[dict]): Transforms to be applied if the condition
            is met. Defaults to [].
        false_transforms (list[dict]): Transforms to be applied if the
            condition is not met. Defaults to [].
    """

    def __init__(self,
                 condition: str,
                 true_transforms: Union[Dict, List[Dict]] = [],
                 false_transforms: Union[Dict, List[Dict]] = []):
        self.condition = condition
        self.true_transforms = Compose(true_transforms)
        self.false_transforms = Compose(false_transforms)

    def transform(self, results: Dict) -> Optional[Dict]:
        """Transform the image.

        Args:
            results (dict):Result dict containing the data to transform.

        Returns:
            dict: Transformed results.
        """
        if eval(self.condition):
            return self.true_transforms(results)  # type: ignore
        else:
            return self.false_transforms(results)

    def __repr__(self):
        repr_str = self.__class__.__name__
        repr_str += f'(condition = {self.condition}, '
        repr_str += f'true_transforms = {self.true_transforms}, '
        repr_str += f'false_transforms = {self.false_transforms})'
        return repr_str