transforms_3d.py 67.1 KB
Newer Older
dingchang's avatar
dingchang committed
1
# Copyright (c) OpenMMLab. All rights reserved.
2
import random
3
import warnings
4
from typing import Dict, List
5
6
7

import cv2
import numpy as np
8
from mmcv import is_tuple_of
9
from mmcv.transforms import BaseTransform
zhangwenwei's avatar
zhangwenwei committed
10

11
from mmdet3d.core import VoxelGenerator
12
13
from mmdet3d.core.bbox import (CameraInstance3DBoxes, DepthInstance3DBoxes,
                               LiDARInstance3DBoxes, box_np_ops)
14
from mmdet3d.core.points import BasePoints
15
from mmdet3d.registry import TRANSFORMS
zhangwenwei's avatar
zhangwenwei committed
16
from mmdet.datasets.pipelines import RandomFlip
zhangwenwei's avatar
zhangwenwei committed
17
18
19
from .data_augment_utils import noise_per_object_v3_


20
@TRANSFORMS.register_module()
21
22
23
24
25
26
27
28
class RandomDropPointsColor(object):
    r"""Randomly set the color of points to all zeros.

    Once this transform is executed, all the points' color will be dropped.
    Refer to `PAConv <https://github.com/CVMI-Lab/PAConv/blob/main/scene_seg/
    util/transform.py#L223>`_ for more details.

    Args:
29
        drop_ratio (float, optional): The probability of dropping point colors.
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
            Defaults to 0.2.
    """

    def __init__(self, drop_ratio=0.2):
        assert isinstance(drop_ratio, (int, float)) and 0 <= drop_ratio <= 1, \
            f'invalid drop_ratio value {drop_ratio}'
        self.drop_ratio = drop_ratio

    def __call__(self, input_dict):
        """Call function to drop point colors.

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
45
            dict: Results after color dropping,
46
47
48
49
50
51
52
                'points' key is updated in the result dict.
        """
        points = input_dict['points']
        assert points.attribute_dims is not None and \
            'color' in points.attribute_dims, \
            'Expect points have color attribute'

53
54
55
56
57
58
59
        # this if-expression is a bit strange
        # `RandomDropPointsColor` is used in training 3D segmentor PAConv
        # we discovered in our experiments that, using
        # `if np.random.rand() > 1.0 - self.drop_ratio` consistently leads to
        # better results than using `if np.random.rand() < self.drop_ratio`
        # so we keep this hack in our codebase
        if np.random.rand() > 1.0 - self.drop_ratio:
60
61
62
63
64
65
66
67
68
69
            points.color = points.color * 0.0
        return input_dict

    def __repr__(self):
        """str: Return a string that describes the module."""
        repr_str = self.__class__.__name__
        repr_str += f'(drop_ratio={self.drop_ratio})'
        return repr_str


70
@TRANSFORMS.register_module()
zhangwenwei's avatar
zhangwenwei committed
71
72
73
74
75
76
77
class RandomFlip3D(RandomFlip):
    """Flip the points & bbox.

    If the input dict contains the key "flip", then the flag will be used,
    otherwise it will be randomly decided by a ratio specified in the init
    method.

jshilong's avatar
jshilong committed
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96

    Required Keys:

    - points (np.float32)
    - gt_bboxes_3d (np.float32)

    Modified Keys:

    - points (np.float32)
    - gt_bboxes_3d (np.float32)

    Added Keys:

    - points (np.float32)
    - pcd_trans (np.float32)
    - pcd_rotation (np.float32)
    - pcd_rotation_angle (np.float32)
    - pcd_scale_factor (np.float32)

zhangwenwei's avatar
zhangwenwei committed
97
    Args:
zhangwenwei's avatar
zhangwenwei committed
98
99
100
        sync_2d (bool, optional): Whether to apply flip according to the 2D
            images. If True, it will apply the same flip as that to 2D images.
            If False, it will decide whether to flip randomly and independently
liyinhao's avatar
liyinhao committed
101
            to that of 2D images. Defaults to True.
wuyuefeng's avatar
wuyuefeng committed
102
        flip_ratio_bev_horizontal (float, optional): The flipping probability
liyinhao's avatar
liyinhao committed
103
            in horizontal direction. Defaults to 0.0.
wuyuefeng's avatar
wuyuefeng committed
104
        flip_ratio_bev_vertical (float, optional): The flipping probability
liyinhao's avatar
liyinhao committed
105
            in vertical direction. Defaults to 0.0.
zhangwenwei's avatar
zhangwenwei committed
106
107
    """

wuyuefeng's avatar
wuyuefeng committed
108
    def __init__(self,
jshilong's avatar
jshilong committed
109
110
111
112
113
114
115
                 sync_2d: bool = True,
                 flip_ratio_bev_horizontal: float = 0.0,
                 flip_ratio_bev_vertical: float = 0.0,
                 **kwargs) -> None:
        # `flip_ratio_bev_horizontal` is equal to
        # for flip prob of 2d image when
        # `sync_2d` is True
wuyuefeng's avatar
wuyuefeng committed
116
        super(RandomFlip3D, self).__init__(
jshilong's avatar
jshilong committed
117
            prob=flip_ratio_bev_horizontal, direction='horizontal', **kwargs)
zhangwenwei's avatar
zhangwenwei committed
118
        self.sync_2d = sync_2d
jshilong's avatar
jshilong committed
119
        self.flip_ratio_bev_horizontal = flip_ratio_bev_horizontal
wuyuefeng's avatar
wuyuefeng committed
120
121
122
123
124
125
126
127
128
129
        self.flip_ratio_bev_vertical = flip_ratio_bev_vertical
        if flip_ratio_bev_horizontal is not None:
            assert isinstance(
                flip_ratio_bev_horizontal,
                (int, float)) and 0 <= flip_ratio_bev_horizontal <= 1
        if flip_ratio_bev_vertical is not None:
            assert isinstance(
                flip_ratio_bev_vertical,
                (int, float)) and 0 <= flip_ratio_bev_vertical <= 1

jshilong's avatar
jshilong committed
130
131
132
    def random_flip_data_3d(self,
                            input_dict: dict,
                            direction: str = 'horizontal') -> None:
133
134
        """Flip 3D data randomly.

jshilong's avatar
jshilong committed
135
136
137
138
139
140
141
        `random_flip_data_3d` should take these situations into consideration:

        - 1. LIDAR-based 3d detection
        - 2. LIDAR-based 3d segmentation
        - 3. vision-only detection
        - 4. multi-modality 3d detection.

142
143
        Args:
            input_dict (dict): Result dict from loading pipeline.
144
145
            direction (str, optional): Flip direction.
                Default: 'horizontal'.
146
147

        Returns:
148
            dict: Flipped results, 'points', 'bbox3d_fields' keys are
149
150
                updated in the result dict.
        """
wuyuefeng's avatar
wuyuefeng committed
151
        assert direction in ['horizontal', 'vertical']
jshilong's avatar
jshilong committed
152
153

        if 'gt_bboxes_3d' in input_dict:
154
            if 'points' in input_dict:
jshilong's avatar
jshilong committed
155
                input_dict['points'] = input_dict['gt_bboxes_3d'].flip(
156
157
                    direction, points=input_dict['points'])
            else:
jshilong's avatar
jshilong committed
158
159
160
161
162
163
                # vision-only detection
                input_dict['gt_bboxes_3d'].flip(direction)
        else:
            input_dict['points'].flip(direction)

        if 'centers_2d' in input_dict:
164
165
            assert self.sync_2d is True and direction == 'horizontal', \
                'Only support sync_2d=True and horizontal flip with images'
jshilong's avatar
jshilong committed
166
167
            # TODO fix this ori_shape and other keys in vision based model
            # TODO ori_shape to img_shape
168
            w = input_dict['ori_shape'][1]
jshilong's avatar
jshilong committed
169
170
            input_dict['centers_2d'][..., 0] = \
                w - input_dict['centers_2d'][..., 0]
171
172
            # need to modify the horizontal position of camera center
            # along u-axis in the image (flip like centers2d)
173
            # ['cam2img'][0][2] = c_u
174
175
            # see more details and examples at
            # https://github.com/open-mmlab/mmdetection3d/pull/744
176
            input_dict['cam2img'][0][2] = w - input_dict['cam2img'][0][2]
zhangwenwei's avatar
zhangwenwei committed
177

jshilong's avatar
jshilong committed
178
    def transform(self, input_dict: dict) -> dict:
179
        """Call function to flip points, values in the ``bbox3d_fields`` and
180
181
182
183
184
185
        also flip 2D image and its annotations.

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
186
187
            dict: Flipped results, 'flip', 'flip_direction',
                'pcd_horizontal_flip' and 'pcd_vertical_flip' keys are added
188
189
                into result dict.
        """
190
        # flip 2D image and its annotations
jshilong's avatar
jshilong committed
191
192
        if 'img' in input_dict:
            super(RandomFlip3D, self).transform(input_dict)
zhangwenwei's avatar
zhangwenwei committed
193

jshilong's avatar
jshilong committed
194
        if self.sync_2d and 'img' in input_dict:
wuyuefeng's avatar
wuyuefeng committed
195
196
            input_dict['pcd_horizontal_flip'] = input_dict['flip']
            input_dict['pcd_vertical_flip'] = False
zhangwenwei's avatar
zhangwenwei committed
197
        else:
wuyuefeng's avatar
wuyuefeng committed
198
199
            if 'pcd_horizontal_flip' not in input_dict:
                flip_horizontal = True if np.random.rand(
jshilong's avatar
jshilong committed
200
                ) < self.flip_ratio_bev_horizontal else False
wuyuefeng's avatar
wuyuefeng committed
201
202
203
204
205
206
                input_dict['pcd_horizontal_flip'] = flip_horizontal
            if 'pcd_vertical_flip' not in input_dict:
                flip_vertical = True if np.random.rand(
                ) < self.flip_ratio_bev_vertical else False
                input_dict['pcd_vertical_flip'] = flip_vertical

207
208
209
        if 'transformation_3d_flow' not in input_dict:
            input_dict['transformation_3d_flow'] = []

wuyuefeng's avatar
wuyuefeng committed
210
211
        if input_dict['pcd_horizontal_flip']:
            self.random_flip_data_3d(input_dict, 'horizontal')
212
            input_dict['transformation_3d_flow'].extend(['HF'])
wuyuefeng's avatar
wuyuefeng committed
213
214
        if input_dict['pcd_vertical_flip']:
            self.random_flip_data_3d(input_dict, 'vertical')
215
            input_dict['transformation_3d_flow'].extend(['VF'])
zhangwenwei's avatar
zhangwenwei committed
216
217
        return input_dict

zhangwenwei's avatar
zhangwenwei committed
218
    def __repr__(self):
219
        """str: Return a string that describes the module."""
wuyuefeng's avatar
wuyuefeng committed
220
        repr_str = self.__class__.__name__
221
        repr_str += f'(sync_2d={self.sync_2d},'
222
        repr_str += f' flip_ratio_bev_vertical={self.flip_ratio_bev_vertical})'
wuyuefeng's avatar
wuyuefeng committed
223
        return repr_str
zhangwenwei's avatar
zhangwenwei committed
224

zhangwenwei's avatar
zhangwenwei committed
225

226
@TRANSFORMS.register_module()
227
228
229
class RandomJitterPoints(object):
    """Randomly jitter point coordinates.

230
    Different from the global translation in ``GlobalRotScaleTrans``, here we
231
232
233
234
        apply different noises to each point in a scene.

    Args:
        jitter_std (list[float]): The standard deviation of jittering noise.
235
236
            This applies random noise to all points in a 3D scene, which is
            sampled from a gaussian distribution whose standard deviation is
237
            set by ``jitter_std``. Defaults to [0.01, 0.01, 0.01]
238
        clip_range (list[float]): Clip the randomly generated jitter
239
240
241
242
            noise into this range. If None is given, don't perform clipping.
            Defaults to [-0.05, 0.05]

    Note:
243
        This transform should only be used in point cloud segmentation tasks
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
            because we don't transform ground-truth bboxes accordingly.
        For similar transform in detection task, please refer to `ObjectNoise`.
    """

    def __init__(self,
                 jitter_std=[0.01, 0.01, 0.01],
                 clip_range=[-0.05, 0.05]):
        seq_types = (list, tuple, np.ndarray)
        if not isinstance(jitter_std, seq_types):
            assert isinstance(jitter_std, (int, float)), \
                f'unsupported jitter_std type {type(jitter_std)}'
            jitter_std = [jitter_std, jitter_std, jitter_std]
        self.jitter_std = jitter_std

        if clip_range is not None:
            if not isinstance(clip_range, seq_types):
                assert isinstance(clip_range, (int, float)), \
                    f'unsupported clip_range type {type(clip_range)}'
                clip_range = [-clip_range, clip_range]
        self.clip_range = clip_range

    def __call__(self, input_dict):
        """Call function to jitter all the points in the scene.

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
272
            dict: Results after adding noise to each point,
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
                'points' key is updated in the result dict.
        """
        points = input_dict['points']
        jitter_std = np.array(self.jitter_std, dtype=np.float32)
        jitter_noise = \
            np.random.randn(points.shape[0], 3) * jitter_std[None, :]
        if self.clip_range is not None:
            jitter_noise = np.clip(jitter_noise, self.clip_range[0],
                                   self.clip_range[1])

        points.translate(jitter_noise)
        return input_dict

    def __repr__(self):
        """str: Return a string that describes the module."""
        repr_str = self.__class__.__name__
        repr_str += f'(jitter_std={self.jitter_std},'
        repr_str += f' clip_range={self.clip_range})'
        return repr_str


294
295
@TRANSFORMS.register_module()
class ObjectSample(BaseTransform):
zhangwenwei's avatar
zhangwenwei committed
296
    """Sample GT objects to the data.
zhangwenwei's avatar
zhangwenwei committed
297

298
299
300
301
302
303
304
305
306
307
    Required Keys:

    - points
    - ann_info
    - gt_bboxes_3d
    - gt_labels_3d
    - img (optional)
    - gt_bboxes (optional)

    Modified Keys:
308

309
310
311
312
313
314
315
316
317
318
    - points
    - gt_bboxes_3d
    - gt_labels_3d
    - img (optional)
    - gt_bboxes (optional)

    Added Keys:

    - plane (optional)

zhangwenwei's avatar
zhangwenwei committed
319
320
321
322
    Args:
        db_sampler (dict): Config dict of the database sampler.
        sample_2d (bool): Whether to also paste 2D image patch to the images
            This should be true when applying multi-modality cut-and-paste.
liyinhao's avatar
liyinhao committed
323
            Defaults to False.
324
        use_ground_plane (bool): Whether to use ground plane to adjust the
325
            3D labels.
zhangwenwei's avatar
zhangwenwei committed
326
    """
zhangwenwei's avatar
zhangwenwei committed
327

328
329
330
331
    def __init__(self,
                 db_sampler: dict,
                 sample_2d: bool = False,
                 use_ground_plane: bool = False):
zhangwenwei's avatar
zhangwenwei committed
332
333
334
335
        self.sampler_cfg = db_sampler
        self.sample_2d = sample_2d
        if 'type' not in db_sampler.keys():
            db_sampler['type'] = 'DataBaseSampler'
336
        self.db_sampler = TRANSFORMS.build(db_sampler)
337
        self.use_ground_plane = use_ground_plane
zhangwenwei's avatar
zhangwenwei committed
338
339

    @staticmethod
340
341
    def remove_points_in_boxes(points: BasePoints,
                               boxes: np.ndarray) -> np.ndarray:
342
343
344
        """Remove the points in the sampled bounding boxes.

        Args:
345
            points (:obj:`BasePoints`): Input point cloud array.
346
347
348
349
350
            boxes (np.ndarray): Sampled ground truth boxes.

        Returns:
            np.ndarray: Points with those in the boxes removed.
        """
351
        masks = box_np_ops.points_in_rbbox(points.coord.numpy(), boxes)
zhangwenwei's avatar
zhangwenwei committed
352
353
354
        points = points[np.logical_not(masks.any(-1))]
        return points

355
356
    def transform(self, input_dict: dict) -> dict:
        """Transform function to sample ground truth objects to the data.
357
358
359
360
361

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
362
363
            dict: Results after object sampling augmentation,
                'points', 'gt_bboxes_3d', 'gt_labels_3d' keys are updated
364
365
                in the result dict.
        """
zhangwenwei's avatar
zhangwenwei committed
366
        gt_bboxes_3d = input_dict['gt_bboxes_3d']
zhangwenwei's avatar
zhangwenwei committed
367
368
        gt_labels_3d = input_dict['gt_labels_3d']

369
370
371
372
373
        if self.use_ground_plane and 'plane' in input_dict['ann_info']:
            ground_plane = input_dict['ann_info']['plane']
            input_dict['plane'] = ground_plane
        else:
            ground_plane = None
zhangwenwei's avatar
zhangwenwei committed
374
375
376
        # change to float for blending operation
        points = input_dict['points']
        if self.sample_2d:
wuyuefeng's avatar
wuyuefeng committed
377
            img = input_dict['img']
zhangwenwei's avatar
zhangwenwei committed
378
379
380
            gt_bboxes_2d = input_dict['gt_bboxes']
            # Assume for now 3D & 2D bboxes are the same
            sampled_dict = self.db_sampler.sample_all(
381
382
383
384
                gt_bboxes_3d.tensor.numpy(),
                gt_labels_3d,
                gt_bboxes_2d=gt_bboxes_2d,
                img=img)
zhangwenwei's avatar
zhangwenwei committed
385
386
        else:
            sampled_dict = self.db_sampler.sample_all(
387
388
389
390
                gt_bboxes_3d.tensor.numpy(),
                gt_labels_3d,
                img=None,
                ground_plane=ground_plane)
zhangwenwei's avatar
zhangwenwei committed
391
392
393
394

        if sampled_dict is not None:
            sampled_gt_bboxes_3d = sampled_dict['gt_bboxes_3d']
            sampled_points = sampled_dict['points']
zhangwenwei's avatar
zhangwenwei committed
395
            sampled_gt_labels = sampled_dict['gt_labels_3d']
zhangwenwei's avatar
zhangwenwei committed
396

zhangwenwei's avatar
zhangwenwei committed
397
398
            gt_labels_3d = np.concatenate([gt_labels_3d, sampled_gt_labels],
                                          axis=0)
399
400
401
            gt_bboxes_3d = gt_bboxes_3d.new_box(
                np.concatenate(
                    [gt_bboxes_3d.tensor.numpy(), sampled_gt_bboxes_3d]))
zhangwenwei's avatar
zhangwenwei committed
402

zhangwenwei's avatar
zhangwenwei committed
403
404
            points = self.remove_points_in_boxes(points, sampled_gt_bboxes_3d)
            # check the points dimension
405
            points = points.cat([sampled_points, points])
zhangwenwei's avatar
zhangwenwei committed
406
407
408
409
410

            if self.sample_2d:
                sampled_gt_bboxes_2d = sampled_dict['gt_bboxes_2d']
                gt_bboxes_2d = np.concatenate(
                    [gt_bboxes_2d, sampled_gt_bboxes_2d]).astype(np.float32)
zhangwenwei's avatar
zhangwenwei committed
411

zhangwenwei's avatar
zhangwenwei committed
412
                input_dict['gt_bboxes'] = gt_bboxes_2d
wuyuefeng's avatar
wuyuefeng committed
413
                input_dict['img'] = sampled_dict['img']
zhangwenwei's avatar
zhangwenwei committed
414
415

        input_dict['gt_bboxes_3d'] = gt_bboxes_3d
WRH's avatar
WRH committed
416
        input_dict['gt_labels_3d'] = gt_labels_3d.astype(np.int64)
zhangwenwei's avatar
zhangwenwei committed
417
        input_dict['points'] = points
zhangwenwei's avatar
zhangwenwei committed
418

zhangwenwei's avatar
zhangwenwei committed
419
420
421
        return input_dict

    def __repr__(self):
422
        """str: Return a string that describes the module."""
423
424
425
426
427
428
429
430
431
        repr_str = self.__class__.__name__
        repr_str += f' sample_2d={self.sample_2d},'
        repr_str += f' data_root={self.sampler_cfg.data_root},'
        repr_str += f' info_path={self.sampler_cfg.info_path},'
        repr_str += f' rate={self.sampler_cfg.rate},'
        repr_str += f' prepare={self.sampler_cfg.prepare},'
        repr_str += f' classes={self.sampler_cfg.classes},'
        repr_str += f' sample_groups={self.sampler_cfg.sample_groups}'
        return repr_str
zhangwenwei's avatar
zhangwenwei committed
432
433


434
435
@TRANSFORMS.register_module()
class ObjectNoise(BaseTransform):
zhangwenwei's avatar
zhangwenwei committed
436
    """Apply noise to each GT objects in the scene.
zhangwenwei's avatar
zhangwenwei committed
437

438
439
440
441
442
443
444
445
446
447
    Required Keys:

    - points
    - gt_bboxes_3d

    Modified Keys:

    - points
    - gt_bboxes_3d

zhangwenwei's avatar
zhangwenwei committed
448
    Args:
449
        translation_std (list[float], optional): Standard deviation of the
zhangwenwei's avatar
zhangwenwei committed
450
451
            distribution where translation noise are sampled from.
            Defaults to [0.25, 0.25, 0.25].
452
        global_rot_range (list[float], optional): Global rotation to the scene.
zhangwenwei's avatar
zhangwenwei committed
453
            Defaults to [0.0, 0.0].
454
        rot_range (list[float], optional): Object rotation range.
zhangwenwei's avatar
zhangwenwei committed
455
456
457
458
            Defaults to [-0.15707963267, 0.15707963267].
        num_try (int, optional): Number of times to try if the noise applied is
            invalid. Defaults to 100.
    """
zhangwenwei's avatar
zhangwenwei committed
459
460

    def __init__(self,
461
462
463
464
                 translation_std: list = [0.25, 0.25, 0.25],
                 global_rot_range: list = [0.0, 0.0],
                 rot_range: list = [-0.15707963267, 0.15707963267],
                 num_try: int = 100):
zhangwenwei's avatar
zhangwenwei committed
465
        self.translation_std = translation_std
zhangwenwei's avatar
zhangwenwei committed
466
        self.global_rot_range = global_rot_range
zhangwenwei's avatar
zhangwenwei committed
467
        self.rot_range = rot_range
zhangwenwei's avatar
zhangwenwei committed
468
469
        self.num_try = num_try

470
471
    def transform(self, input_dict: dict) -> dict:
        """Transform function to apply noise to each ground truth in the scene.
472
473
474
475
476

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
477
            dict: Results after adding noise to each object,
478
479
                'points', 'gt_bboxes_3d' keys are updated in the result dict.
        """
zhangwenwei's avatar
zhangwenwei committed
480
481
        gt_bboxes_3d = input_dict['gt_bboxes_3d']
        points = input_dict['points']
zhangwenwei's avatar
zhangwenwei committed
482

483
        # TODO: this is inplace operation
484
        numpy_box = gt_bboxes_3d.tensor.numpy()
485
486
        numpy_points = points.tensor.numpy()

zhangwenwei's avatar
zhangwenwei committed
487
        noise_per_object_v3_(
488
            numpy_box,
489
            numpy_points,
zhangwenwei's avatar
zhangwenwei committed
490
491
            rotation_perturb=self.rot_range,
            center_noise_std=self.translation_std,
zhangwenwei's avatar
zhangwenwei committed
492
493
            global_random_rot_range=self.global_rot_range,
            num_try=self.num_try)
494
495

        input_dict['gt_bboxes_3d'] = gt_bboxes_3d.new_box(numpy_box)
496
        input_dict['points'] = points.new_point(numpy_points)
zhangwenwei's avatar
zhangwenwei committed
497
498
499
        return input_dict

    def __repr__(self):
500
        """str: Return a string that describes the module."""
zhangwenwei's avatar
zhangwenwei committed
501
        repr_str = self.__class__.__name__
502
503
504
505
        repr_str += f'(num_try={self.num_try},'
        repr_str += f' translation_std={self.translation_std},'
        repr_str += f' global_rot_range={self.global_rot_range},'
        repr_str += f' rot_range={self.rot_range})'
zhangwenwei's avatar
zhangwenwei committed
506
507
508
        return repr_str


509
@TRANSFORMS.register_module()
510
class GlobalAlignment(BaseTransform):
511
512
513
514
515
516
    """Apply global alignment to 3D scene points by rotation and translation.

    Args:
        rotation_axis (int): Rotation axis for points and bboxes rotation.

    Note:
517
518
        We do not record the applied rotation and translation as in
            GlobalRotScaleTrans. Because usually, we do not need to reverse
519
            the alignment step.
520
        For example, ScanNet 3D detection task uses aligned ground-truth
521
522
523
            bounding boxes for evaluation.
    """

524
    def __init__(self, rotation_axis: int) -> None:
525
526
        self.rotation_axis = rotation_axis

527
    def _trans_points(self, results: Dict, trans_factor: np.ndarray) -> None:
528
529
530
531
532
533
534
535
536
        """Private function to translate points.

        Args:
            input_dict (dict): Result dict from loading pipeline.
            trans_factor (np.ndarray): Translation vector to be applied.

        Returns:
            dict: Results after translation, 'points' is updated in the dict.
        """
537
        results['points'].translate(trans_factor)
538

539
    def _rot_points(self, results: Dict, rot_mat: np.ndarray) -> None:
540
541
542
543
544
545
546
547
548
549
        """Private function to rotate bounding boxes and points.

        Args:
            input_dict (dict): Result dict from loading pipeline.
            rot_mat (np.ndarray): Rotation matrix to be applied.

        Returns:
            dict: Results after rotation, 'points' is updated in the dict.
        """
        # input should be rot_mat_T so I transpose it here
550
        results['points'].rotate(rot_mat.T)
551

552
    def _check_rot_mat(self, rot_mat: np.ndarray) -> None:
553
554
555
556
557
558
559
560
561
562
563
564
        """Check if rotation matrix is valid for self.rotation_axis.

        Args:
            rot_mat (np.ndarray): Rotation matrix to be applied.
        """
        is_valid = np.allclose(np.linalg.det(rot_mat), 1.0)
        valid_array = np.zeros(3)
        valid_array[self.rotation_axis] = 1.0
        is_valid &= (rot_mat[self.rotation_axis, :] == valid_array).all()
        is_valid &= (rot_mat[:, self.rotation_axis] == valid_array).all()
        assert is_valid, f'invalid rotation matrix {rot_mat}'

565
    def transform(self, results: Dict) -> Dict:
566
567
568
569
570
571
        """Call function to shuffle points.

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
572
            dict: Results after global alignment, 'points' and keys in
573
574
                input_dict['bbox3d_fields'] are updated in the result dict.
        """
575
        assert 'axis_align_matrix' in results, \
576
577
            'axis_align_matrix is not provided in GlobalAlignment'

578
        axis_align_matrix = results['axis_align_matrix']
579
580
581
582
583
584
        assert axis_align_matrix.shape == (4, 4), \
            f'invalid shape {axis_align_matrix.shape} for axis_align_matrix'
        rot_mat = axis_align_matrix[:3, :3]
        trans_vec = axis_align_matrix[:3, -1]

        self._check_rot_mat(rot_mat)
585
586
        self._rot_points(results, rot_mat)
        self._trans_points(results, trans_vec)
587

588
        return results
589
590
591
592
593
594
595

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


596
@TRANSFORMS.register_module()
jshilong's avatar
jshilong committed
597
class GlobalRotScaleTrans(BaseTransform):
zhangwenwei's avatar
zhangwenwei committed
598
    """Apply global rotation, scaling and translation to a 3D scene.
zhangwenwei's avatar
zhangwenwei committed
599

jshilong's avatar
jshilong committed
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
    Required Keys:

    - points (np.float32)
    - gt_bboxes_3d (np.float32)

    Modified Keys:

    - points (np.float32)
    - gt_bboxes_3d (np.float32)

    Added Keys:

    - points (np.float32)
    - pcd_trans (np.float32)
    - pcd_rotation (np.float32)
    - pcd_rotation_angle (np.float32)
    - pcd_scale_factor (np.float32)

zhangwenwei's avatar
zhangwenwei committed
618
    Args:
619
        rot_range (list[float], optional): Range of rotation angle.
liyinhao's avatar
liyinhao committed
620
            Defaults to [-0.78539816, 0.78539816] (close to [-pi/4, pi/4]).
621
        scale_ratio_range (list[float], optional): Range of scale ratio.
liyinhao's avatar
liyinhao committed
622
            Defaults to [0.95, 1.05].
623
624
        translation_std (list[float], optional): The standard deviation of
            translation noise applied to a scene, which
zhangwenwei's avatar
zhangwenwei committed
625
            is sampled from a gaussian distribution whose standard deviation
liyinhao's avatar
liyinhao committed
626
            is set by ``translation_std``. Defaults to [0, 0, 0]
627
        shift_height (bool, optional): Whether to shift height.
wuyuefeng's avatar
wuyuefeng committed
628
            (the fourth dimension of indoor points) when scaling.
liyinhao's avatar
liyinhao committed
629
            Defaults to False.
zhangwenwei's avatar
zhangwenwei committed
630
    """
zhangwenwei's avatar
zhangwenwei committed
631
632

    def __init__(self,
jshilong's avatar
jshilong committed
633
634
635
636
                 rot_range: List[float] = [-0.78539816, 0.78539816],
                 scale_ratio_range: List[float] = [0.95, 1.05],
                 translation_std: List[int] = [0, 0, 0],
                 shift_height: bool = False) -> None:
637
638
639
640
641
        seq_types = (list, tuple, np.ndarray)
        if not isinstance(rot_range, seq_types):
            assert isinstance(rot_range, (int, float)), \
                f'unsupported rot_range type {type(rot_range)}'
            rot_range = [-rot_range, rot_range]
zhangwenwei's avatar
zhangwenwei committed
642
        self.rot_range = rot_range
643
644
645

        assert isinstance(scale_ratio_range, seq_types), \
            f'unsupported scale_ratio_range type {type(scale_ratio_range)}'
jshilong's avatar
jshilong committed
646

zhangwenwei's avatar
zhangwenwei committed
647
        self.scale_ratio_range = scale_ratio_range
648
649
650
651
652
653
654

        if not isinstance(translation_std, seq_types):
            assert isinstance(translation_std, (int, float)), \
                f'unsupported translation_std type {type(translation_std)}'
            translation_std = [
                translation_std, translation_std, translation_std
            ]
655
656
        assert all([std >= 0 for std in translation_std]), \
            'translation_std should be positive'
zhangwenwei's avatar
zhangwenwei committed
657
        self.translation_std = translation_std
wuyuefeng's avatar
wuyuefeng committed
658
        self.shift_height = shift_height
zhangwenwei's avatar
zhangwenwei committed
659

jshilong's avatar
jshilong committed
660
    def _trans_bbox_points(self, input_dict: dict) -> None:
661
662
663
664
665
666
        """Private function to translate bounding boxes and points.

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
667
            dict: Results after translation, 'points', 'pcd_trans'
jshilong's avatar
jshilong committed
668
669
            and `gt_bboxes_3d` is updated
            in the result dict.
670
        """
671
        translation_std = np.array(self.translation_std, dtype=np.float32)
zhangwenwei's avatar
zhangwenwei committed
672
673
        trans_factor = np.random.normal(scale=translation_std, size=3).T

674
        input_dict['points'].translate(trans_factor)
zhangwenwei's avatar
zhangwenwei committed
675
        input_dict['pcd_trans'] = trans_factor
jshilong's avatar
jshilong committed
676
677
        if 'gt_bboxes_3d' in input_dict:
            input_dict['gt_bboxes_3d'].translate(trans_factor)
zhangwenwei's avatar
zhangwenwei committed
678

jshilong's avatar
jshilong committed
679
    def _rot_bbox_points(self, input_dict: dict) -> None:
680
681
682
683
684
685
        """Private function to rotate bounding boxes and points.

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
686
            dict: Results after rotation, 'points', 'pcd_rotation'
jshilong's avatar
jshilong committed
687
688
            and `gt_bboxes_3d` is updated
            in the result dict.
689
        """
zhangwenwei's avatar
zhangwenwei committed
690
        rotation = self.rot_range
zhangwenwei's avatar
zhangwenwei committed
691
        noise_rotation = np.random.uniform(rotation[0], rotation[1])
zhangwenwei's avatar
zhangwenwei committed
692

jshilong's avatar
jshilong committed
693
694
695
696
697
698
699
700
        if 'gt_bboxes_3d' in input_dict and \
                len(input_dict['gt_bboxes_3d'].tensor) != 0:
            # rotate points with bboxes
            points, rot_mat_T = input_dict['gt_bboxes_3d'].rotate(
                noise_rotation, input_dict['points'])
            input_dict['points'] = points
        else:
            # if no bbox in input_dict, only rotate points
701
            rot_mat_T = input_dict['points'].rotate(noise_rotation)
jshilong's avatar
jshilong committed
702
703
704
705
706

        input_dict['pcd_rotation'] = rot_mat_T
        input_dict['pcd_rotation_angle'] = noise_rotation

    def _scale_bbox_points(self, input_dict: dict) -> None:
707
708
709
710
711
712
        """Private function to scale bounding boxes and points.

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
jshilong's avatar
jshilong committed
713
714
715
            dict: Results after scaling, 'points' and
            `gt_bboxes_3d` is updated
            in the result dict.
716
        """
zhangwenwei's avatar
zhangwenwei committed
717
        scale = input_dict['pcd_scale_factor']
718
719
        points = input_dict['points']
        points.scale(scale)
wuyuefeng's avatar
wuyuefeng committed
720
        if self.shift_height:
721
722
            assert 'height' in points.attribute_dims.keys(), \
                'setting shift_height=True but points have no height attribute'
723
724
            points.tensor[:, points.attribute_dims['height']] *= scale
        input_dict['points'] = points
wuyuefeng's avatar
wuyuefeng committed
725

jshilong's avatar
jshilong committed
726
727
728
        if 'gt_bboxes_3d' in input_dict and \
                len(input_dict['gt_bboxes_3d'].tensor) != 0:
            input_dict['gt_bboxes_3d'].scale(scale)
zhangwenwei's avatar
zhangwenwei committed
729

jshilong's avatar
jshilong committed
730
    def _random_scale(self, input_dict: dict) -> None:
731
732
733
734
735
736
        """Private function to randomly set the scale factor.

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
jshilong's avatar
jshilong committed
737
738
            dict: Results after scaling, 'pcd_scale_factor'
            are updated in the result dict.
739
        """
zhangwenwei's avatar
zhangwenwei committed
740
741
742
        scale_factor = np.random.uniform(self.scale_ratio_range[0],
                                         self.scale_ratio_range[1])
        input_dict['pcd_scale_factor'] = scale_factor
zhangwenwei's avatar
zhangwenwei committed
743

jshilong's avatar
jshilong committed
744
    def transform(self, input_dict: dict) -> dict:
745
        """Private function to rotate, scale and translate bounding boxes and
746
747
748
749
750
751
752
        points.

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
            dict: Results after scaling, 'points', 'pcd_rotation',
jshilong's avatar
jshilong committed
753
754
            'pcd_scale_factor', 'pcd_trans' and `gt_bboxes_3d` is updated
            in the result dict.
755
        """
756
757
758
        if 'transformation_3d_flow' not in input_dict:
            input_dict['transformation_3d_flow'] = []

zhangwenwei's avatar
zhangwenwei committed
759
        self._rot_bbox_points(input_dict)
zhangwenwei's avatar
zhangwenwei committed
760

zhangwenwei's avatar
zhangwenwei committed
761
762
763
        if 'pcd_scale_factor' not in input_dict:
            self._random_scale(input_dict)
        self._scale_bbox_points(input_dict)
zhangwenwei's avatar
zhangwenwei committed
764

zhangwenwei's avatar
zhangwenwei committed
765
        self._trans_bbox_points(input_dict)
766
767

        input_dict['transformation_3d_flow'].extend(['R', 'S', 'T'])
zhangwenwei's avatar
zhangwenwei committed
768
769
770
        return input_dict

    def __repr__(self):
771
        """str: Return a string that describes the module."""
zhangwenwei's avatar
zhangwenwei committed
772
        repr_str = self.__class__.__name__
773
774
775
776
        repr_str += f'(rot_range={self.rot_range},'
        repr_str += f' scale_ratio_range={self.scale_ratio_range},'
        repr_str += f' translation_std={self.translation_std},'
        repr_str += f' shift_height={self.shift_height})'
zhangwenwei's avatar
zhangwenwei committed
777
778
779
        return repr_str


780
@TRANSFORMS.register_module()
zhangwenwei's avatar
zhangwenwei committed
781
class PointShuffle(object):
782
    """Shuffle input points."""
zhangwenwei's avatar
zhangwenwei committed
783
784

    def __call__(self, input_dict):
785
786
787
788
789
790
        """Call function to shuffle points.

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
791
            dict: Results after filtering, 'points', 'pts_instance_mask'
792
                and 'pts_semantic_mask' keys are updated in the result dict.
793
        """
794
795
796
797
798
799
800
801
802
803
804
805
        idx = input_dict['points'].shuffle()
        idx = idx.numpy()

        pts_instance_mask = input_dict.get('pts_instance_mask', None)
        pts_semantic_mask = input_dict.get('pts_semantic_mask', None)

        if pts_instance_mask is not None:
            input_dict['pts_instance_mask'] = pts_instance_mask[idx]

        if pts_semantic_mask is not None:
            input_dict['pts_semantic_mask'] = pts_semantic_mask[idx]

zhangwenwei's avatar
zhangwenwei committed
806
807
808
809
810
811
        return input_dict

    def __repr__(self):
        return self.__class__.__name__


812
@TRANSFORMS.register_module()
813
class ObjectRangeFilter(BaseTransform):
814
815
    """Filter objects by the range.

816
817
818
819
820
821
822
823
    Required Keys:

    - gt_bboxes_3d

    Modified Keys:

    - gt_bboxes_3d

824
825
826
    Args:
        point_cloud_range (list[float]): Point cloud range.
    """
zhangwenwei's avatar
zhangwenwei committed
827

828
    def __init__(self, point_cloud_range: list):
zhangwenwei's avatar
zhangwenwei committed
829
830
        self.pcd_range = np.array(point_cloud_range, dtype=np.float32)

831
832
    def transform(self, input_dict: dict) -> dict:
        """Transform function to filter objects by the range.
833
834
835
836
837

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
838
            dict: Results after filtering, 'gt_bboxes_3d', 'gt_labels_3d'
839
840
                keys are updated in the result dict.
        """
841
842
843
844
845
846
847
        # Check points instance type and initialise bev_range
        if isinstance(input_dict['gt_bboxes_3d'],
                      (LiDARInstance3DBoxes, DepthInstance3DBoxes)):
            bev_range = self.pcd_range[[0, 1, 3, 4]]
        elif isinstance(input_dict['gt_bboxes_3d'], CameraInstance3DBoxes):
            bev_range = self.pcd_range[[0, 2, 3, 5]]

zhangwenwei's avatar
zhangwenwei committed
848
        gt_bboxes_3d = input_dict['gt_bboxes_3d']
zhangwenwei's avatar
zhangwenwei committed
849
        gt_labels_3d = input_dict['gt_labels_3d']
850
        mask = gt_bboxes_3d.in_range_bev(bev_range)
zhangwenwei's avatar
zhangwenwei committed
851
        gt_bboxes_3d = gt_bboxes_3d[mask]
ZwwWayne's avatar
ZwwWayne committed
852
853
854
855
856
        # mask is a torch tensor but gt_labels_3d is still numpy array
        # using mask to index gt_labels_3d will cause bug when
        # len(gt_labels_3d) == 1, where mask=1 will be interpreted
        # as gt_labels_3d[1] and cause out of index error
        gt_labels_3d = gt_labels_3d[mask.numpy().astype(np.bool)]
zhangwenwei's avatar
zhangwenwei committed
857
858

        # limit rad to [-pi, pi]
859
860
        gt_bboxes_3d.limit_yaw(offset=0.5, period=2 * np.pi)
        input_dict['gt_bboxes_3d'] = gt_bboxes_3d
zhangwenwei's avatar
zhangwenwei committed
861
862
        input_dict['gt_labels_3d'] = gt_labels_3d

zhangwenwei's avatar
zhangwenwei committed
863
864
865
        return input_dict

    def __repr__(self):
866
        """str: Return a string that describes the module."""
zhangwenwei's avatar
zhangwenwei committed
867
        repr_str = self.__class__.__name__
868
        repr_str += f'(point_cloud_range={self.pcd_range.tolist()})'
zhangwenwei's avatar
zhangwenwei committed
869
870
871
        return repr_str


872
@TRANSFORMS.register_module()
873
class PointsRangeFilter(BaseTransform):
874
875
    """Filter points by the range.

876
877
878
879
880
881
882
883
884
885
    Required Keys:

    - points
    - pts_instance_mask (optional)

    Modified Keys:

    - points
    - pts_instance_mask (optional)

886
887
888
    Args:
        point_cloud_range (list[float]): Point cloud range.
    """
zhangwenwei's avatar
zhangwenwei committed
889

890
    def __init__(self, point_cloud_range: list):
891
        self.pcd_range = np.array(point_cloud_range, dtype=np.float32)
zhangwenwei's avatar
zhangwenwei committed
892

893
894
    def transform(self, input_dict: dict) -> dict:
        """Transform function to filter points by the range.
895
896
897
898
899

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
900
            dict: Results after filtering, 'points', 'pts_instance_mask'
901
                and 'pts_semantic_mask' keys are updated in the result dict.
902
        """
zhangwenwei's avatar
zhangwenwei committed
903
        points = input_dict['points']
904
905
        points_mask = points.in_range_3d(self.pcd_range)
        clean_points = points[points_mask]
zhangwenwei's avatar
zhangwenwei committed
906
        input_dict['points'] = clean_points
907
908
909
910
911
912
913
914
915
916
917
        points_mask = points_mask.numpy()

        pts_instance_mask = input_dict.get('pts_instance_mask', None)
        pts_semantic_mask = input_dict.get('pts_semantic_mask', None)

        if pts_instance_mask is not None:
            input_dict['pts_instance_mask'] = pts_instance_mask[points_mask]

        if pts_semantic_mask is not None:
            input_dict['pts_semantic_mask'] = pts_semantic_mask[points_mask]

zhangwenwei's avatar
zhangwenwei committed
918
919
920
        return input_dict

    def __repr__(self):
921
        """str: Return a string that describes the module."""
zhangwenwei's avatar
zhangwenwei committed
922
        repr_str = self.__class__.__name__
923
        repr_str += f'(point_cloud_range={self.pcd_range.tolist()})'
zhangwenwei's avatar
zhangwenwei committed
924
        return repr_str
zhangwenwei's avatar
zhangwenwei committed
925
926


927
@TRANSFORMS.register_module()
928
class ObjectNameFilter(BaseTransform):
zhangwenwei's avatar
zhangwenwei committed
929
    """Filter GT objects by their names.
zhangwenwei's avatar
zhangwenwei committed
930

931
932
933
934
935
936
937
938
    Required Keys:

    - gt_labels_3d

    Modified Keys:

    - gt_labels_3d

zhangwenwei's avatar
zhangwenwei committed
939
    Args:
liyinhao's avatar
liyinhao committed
940
        classes (list[str]): List of class names to be kept for training.
zhangwenwei's avatar
zhangwenwei committed
941
942
    """

943
    def __init__(self, classes: list):
zhangwenwei's avatar
zhangwenwei committed
944
945
946
        self.classes = classes
        self.labels = list(range(len(self.classes)))

947
948
    def transform(self, input_dict: dict) -> dict:
        """Transform function to filter objects by their names.
949
950
951
952
953

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
954
            dict: Results after filtering, 'gt_bboxes_3d', 'gt_labels_3d'
955
956
                keys are updated in the result dict.
        """
zhangwenwei's avatar
zhangwenwei committed
957
958
959
960
961
962
963
964
965
        gt_labels_3d = input_dict['gt_labels_3d']
        gt_bboxes_mask = np.array([n in self.labels for n in gt_labels_3d],
                                  dtype=np.bool_)
        input_dict['gt_bboxes_3d'] = input_dict['gt_bboxes_3d'][gt_bboxes_mask]
        input_dict['gt_labels_3d'] = input_dict['gt_labels_3d'][gt_bboxes_mask]

        return input_dict

    def __repr__(self):
966
        """str: Return a string that describes the module."""
zhangwenwei's avatar
zhangwenwei committed
967
968
969
        repr_str = self.__class__.__name__
        repr_str += f'(classes={self.classes})'
        return repr_str
wuyuefeng's avatar
wuyuefeng committed
970
971


972
973
@TRANSFORMS.register_module()
class PointSample(BaseTransform):
974
    """Point sample.
wuyuefeng's avatar
wuyuefeng committed
975
976
977

    Sampling data to a certain number.

978
    Required Keys:
979

980
981
982
983
984
    - points
    - pts_instance_mask (optional)
    - pts_semantic_mask (optional)

    Modified Keys:
985

986
987
988
989
    - points
    - pts_instance_mask (optional)
    - pts_semantic_mask (optional)

wuyuefeng's avatar
wuyuefeng committed
990
991
    Args:
        num_points (int): Number of points to be sampled.
992
        sample_range (float, optional): The range where to sample points.
993
994
995
996
            If not None, the points with depth larger than `sample_range` are
            prior to be sampled. Defaults to None.
        replace (bool, optional): Whether the sampling is with or without
            replacement. Defaults to False.
wuyuefeng's avatar
wuyuefeng committed
997
998
    """

999
1000
1001
1002
    def __init__(self,
                 num_points: int,
                 sample_range: float = None,
                 replace: bool = False):
wuyuefeng's avatar
wuyuefeng committed
1003
        self.num_points = num_points
1004
1005
1006
1007
1008
1009
1010
1011
1012
        self.sample_range = sample_range
        self.replace = replace

    def _points_random_sampling(self,
                                points,
                                num_samples,
                                sample_range=None,
                                replace=False,
                                return_choices=False):
wuyuefeng's avatar
wuyuefeng committed
1013
1014
1015
1016
1017
        """Points random sampling.

        Sample points to a certain number.

        Args:
1018
            points (np.ndarray | :obj:`BasePoints`): 3D Points.
wuyuefeng's avatar
wuyuefeng committed
1019
            num_samples (int): Number of samples to be sampled.
1020
            sample_range (float, optional): Indicating the range where the
1021
                points will be sampled. Defaults to None.
1022
1023
1024
1025
            replace (bool, optional): Sampling with or without replacement.
                Defaults to None.
            return_choices (bool, optional): Whether return choice.
                Defaults to False.
wuyuefeng's avatar
wuyuefeng committed
1026
        Returns:
1027
            tuple[np.ndarray] | np.ndarray:
1028
                - points (np.ndarray | :obj:`BasePoints`): 3D Points.
1029
                - choices (np.ndarray, optional): The generated random samples.
wuyuefeng's avatar
wuyuefeng committed
1030
        """
1031
        if not replace:
wuyuefeng's avatar
wuyuefeng committed
1032
            replace = (points.shape[0] < num_samples)
1033
1034
1035
        point_range = range(len(points))
        if sample_range is not None and not replace:
            # Only sampling the near points when len(points) >= num_samples
1036
1037
1038
            dist = np.linalg.norm(points.tensor, axis=1)
            far_inds = np.where(dist >= sample_range)[0]
            near_inds = np.where(dist < sample_range)[0]
1039
1040
1041
1042
            # in case there are too many far points
            if len(far_inds) > num_samples:
                far_inds = np.random.choice(
                    far_inds, num_samples, replace=False)
1043
1044
1045
1046
1047
1048
1049
            point_range = near_inds
            num_samples -= len(far_inds)
        choices = np.random.choice(point_range, num_samples, replace=replace)
        if sample_range is not None and not replace:
            choices = np.concatenate((far_inds, choices))
            # Shuffle points after sampling
            np.random.shuffle(choices)
wuyuefeng's avatar
wuyuefeng committed
1050
1051
1052
1053
1054
        if return_choices:
            return points[choices], choices
        else:
            return points[choices]

1055
    def transform(self, input_dict: dict) -> dict:
1056
        """Transform function to sample points to in indoor scenes.
1057
1058
1059
1060

        Args:
            input_dict (dict): Result dict from loading pipeline.
        Returns:
1061
            dict: Results after sampling, 'points', 'pts_instance_mask'
1062
1063
                and 'pts_semantic_mask' keys are updated in the result dict.
        """
1064
        points = input_dict['points']
1065
1066
1067
1068
1069
1070
        points, choices = self._points_random_sampling(
            points,
            self.num_points,
            self.sample_range,
            self.replace,
            return_choices=True)
1071
        input_dict['points'] = points
1072

1073
1074
        pts_instance_mask = input_dict.get('pts_instance_mask', None)
        pts_semantic_mask = input_dict.get('pts_semantic_mask', None)
wuyuefeng's avatar
wuyuefeng committed
1075

1076
        if pts_instance_mask is not None:
wuyuefeng's avatar
wuyuefeng committed
1077
            pts_instance_mask = pts_instance_mask[choices]
1078
            input_dict['pts_instance_mask'] = pts_instance_mask
1079
1080
1081

        if pts_semantic_mask is not None:
            pts_semantic_mask = pts_semantic_mask[choices]
1082
            input_dict['pts_semantic_mask'] = pts_semantic_mask
wuyuefeng's avatar
wuyuefeng committed
1083

1084
        return input_dict
wuyuefeng's avatar
wuyuefeng committed
1085
1086

    def __repr__(self):
1087
        """str: Return a string that describes the module."""
wuyuefeng's avatar
wuyuefeng committed
1088
        repr_str = self.__class__.__name__
1089
        repr_str += f'(num_points={self.num_points},'
1090
1091
        repr_str += f' sample_range={self.sample_range},'
        repr_str += f' replace={self.replace})'
1092

1093
1094
1095
        return repr_str


1096
@TRANSFORMS.register_module()
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
class IndoorPointSample(PointSample):
    """Indoor point sample.

    Sampling data to a certain number.
    NOTE: IndoorPointSample is deprecated in favor of PointSample

    Args:
        num_points (int): Number of points to be sampled.
    """

    def __init__(self, *args, **kwargs):
        warnings.warn(
            'IndoorPointSample is deprecated in favor of PointSample')
        super(IndoorPointSample, self).__init__(*args, **kwargs)


1113
@TRANSFORMS.register_module()
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
class IndoorPatchPointSample(object):
    r"""Indoor point sample within a patch. Modified from `PointNet++ <https://
    github.com/charlesq34/pointnet2/blob/master/scannet/scannet_dataset.py>`_.

    Sampling data to a certain number for semantic segmentation.

    Args:
        num_points (int): Number of points to be sampled.
        block_size (float, optional): Size of a block to sample points from.
            Defaults to 1.5.
        sample_rate (float, optional): Stride used in sliding patch generation.
1125
1126
1127
            This parameter is unused in `IndoorPatchPointSample` and thus has
            been deprecated. We plan to remove it in the future.
            Defaults to None.
1128
1129
        ignore_index (int, optional): Label index that won't be used for the
            segmentation task. This is set in PointSegClassMapping as neg_cls.
1130
            If not None, will be used as a patch selection criterion.
1131
1132
1133
1134
1135
            Defaults to None.
        use_normalized_coord (bool, optional): Whether to use normalized xyz as
            additional features. Defaults to False.
        num_try (int, optional): Number of times to try if the patch selected
            is invalid. Defaults to 10.
1136
        enlarge_size (float, optional): Enlarge the sampled patch to
1137
            [-block_size / 2 - enlarge_size, block_size / 2 + enlarge_size] as
1138
            an augmentation. If None, set it as 0. Defaults to 0.2.
1139
        min_unique_num (int, optional): Minimum number of unique points
1140
1141
            the sampled patch should contain. If None, use PointNet++'s method
            to judge uniqueness. Defaults to None.
1142
1143
        eps (float, optional): A value added to patch boundary to guarantee
            points coverage. Defaults to 1e-2.
1144
1145
1146
1147
1148
1149

    Note:
        This transform should only be used in the training process of point
            cloud segmentation tasks. For the sliding patch generation and
            inference process in testing, please refer to the `slide_inference`
            function of `EncoderDecoder3D` class.
1150
1151
1152
1153
1154
    """

    def __init__(self,
                 num_points,
                 block_size=1.5,
1155
                 sample_rate=None,
1156
1157
                 ignore_index=None,
                 use_normalized_coord=False,
1158
1159
                 num_try=10,
                 enlarge_size=0.2,
1160
1161
                 min_unique_num=None,
                 eps=1e-2):
1162
1163
1164
1165
1166
        self.num_points = num_points
        self.block_size = block_size
        self.ignore_index = ignore_index
        self.use_normalized_coord = use_normalized_coord
        self.num_try = num_try
1167
        self.enlarge_size = enlarge_size if enlarge_size is not None else 0.0
1168
        self.min_unique_num = min_unique_num
1169
        self.eps = eps
1170
1171
1172
1173
1174

        if sample_rate is not None:
            warnings.warn(
                "'sample_rate' has been deprecated and will be removed in "
                'the future. Please remove them from your code.')
1175
1176
1177
1178
1179

    def _input_generation(self, coords, patch_center, coord_max, attributes,
                          attribute_dims, point_type):
        """Generating model input.

1180
        Generate input by subtracting patch center and adding additional
1181
1182
1183
1184
1185
1186
1187
1188
1189
            features. Currently support colors and normalized xyz as features.

        Args:
            coords (np.ndarray): Sampled 3D Points.
            patch_center (np.ndarray): Center coordinate of the selected patch.
            coord_max (np.ndarray): Max coordinate of all 3D Points.
            attributes (np.ndarray): features of input points.
            attribute_dims (dict): Dictionary to indicate the meaning of extra
                dimension.
1190
            point_type (type): class of input points inherited from BasePoints.
1191
1192

        Returns:
1193
            :obj:`BasePoints`: The generated input data.
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
        """
        # subtract patch center, the z dimension is not centered
        centered_coords = coords.copy()
        centered_coords[:, 0] -= patch_center[0]
        centered_coords[:, 1] -= patch_center[1]

        if self.use_normalized_coord:
            normalized_coord = coords / coord_max
            attributes = np.concatenate([attributes, normalized_coord], axis=1)
            if attribute_dims is None:
                attribute_dims = dict()
            attribute_dims.update(
                dict(normalized_coord=[
                    attributes.shape[1], attributes.shape[1] +
                    1, attributes.shape[1] + 2
                ]))

        points = np.concatenate([centered_coords, attributes], axis=1)
        points = point_type(
            points, points_dim=points.shape[1], attribute_dims=attribute_dims)

        return points

1217
    def _patch_points_sampling(self, points, sem_mask):
1218
1219
1220
1221
1222
1223
        """Patch points sampling.

        First sample a valid patch.
        Then sample points within that patch to a certain number.

        Args:
1224
            points (:obj:`BasePoints`): 3D Points.
1225
1226
1227
            sem_mask (np.ndarray): semantic segmentation mask for input points.

        Returns:
1228
            tuple[:obj:`BasePoints`, np.ndarray] | :obj:`BasePoints`:
1229

1230
                - points (:obj:`BasePoints`): 3D Points.
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
                - choices (np.ndarray): The generated random samples.
        """
        coords = points.coord.numpy()
        attributes = points.tensor[:, 3:].numpy()
        attribute_dims = points.attribute_dims
        point_type = type(points)

        coord_max = np.amax(coords, axis=0)
        coord_min = np.amin(coords, axis=0)

1241
        for _ in range(self.num_try):
1242
1243
1244
            # random sample a point as patch center
            cur_center = coords[np.random.choice(coords.shape[0])]

1245
1246
            # boundary of a patch, which would be enlarged by
            # `self.enlarge_size` as an augmentation
1247
1248
1249
1250
1251
1252
1253
            cur_max = cur_center + np.array(
                [self.block_size / 2.0, self.block_size / 2.0, 0.0])
            cur_min = cur_center - np.array(
                [self.block_size / 2.0, self.block_size / 2.0, 0.0])
            cur_max[2] = coord_max[2]
            cur_min[2] = coord_min[2]
            cur_choice = np.sum(
1254
1255
                (coords >= (cur_min - self.enlarge_size)) *
                (coords <= (cur_max + self.enlarge_size)),
1256
1257
1258
1259
1260
1261
1262
                axis=1) == 3

            if not cur_choice.any():  # no points in this patch
                continue

            cur_coords = coords[cur_choice, :]
            cur_sem_mask = sem_mask[cur_choice]
1263
            point_idxs = np.where(cur_choice)[0]
1264
            mask = np.sum(
1265
1266
                (cur_coords >= (cur_min - self.eps)) * (cur_coords <=
                                                        (cur_max + self.eps)),
1267
                axis=1) == 3
1268

1269
1270
            # two criteria for patch sampling, adopted from PointNet++
            # 1. selected patch should contain enough unique points
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
            if self.min_unique_num is None:
                # use PointNet++'s method as default
                # [31, 31, 62] are just some big values used to transform
                # coords from 3d array to 1d and then check their uniqueness
                # this is used in all the ScanNet code following PointNet++
                vidx = np.ceil(
                    (cur_coords[mask, :] - cur_min) / (cur_max - cur_min) *
                    np.array([31.0, 31.0, 62.0]))
                vidx = np.unique(vidx[:, 0] * 31.0 * 62.0 + vidx[:, 1] * 62.0 +
                                 vidx[:, 2])
                flag1 = len(vidx) / 31.0 / 31.0 / 62.0 >= 0.02
            else:
1283
                # if `min_unique_num` is provided, directly compare with it
1284
                flag1 = mask.sum() >= self.min_unique_num
1285

1286
            # 2. selected patch should contain enough annotated points
1287
1288
1289
1290
1291
1292
1293
1294
1295
            if self.ignore_index is None:
                flag2 = True
            else:
                flag2 = np.sum(cur_sem_mask != self.ignore_index) / \
                               len(cur_sem_mask) >= 0.7

            if flag1 and flag2:
                break

1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
        # sample idx to `self.num_points`
        if point_idxs.size >= self.num_points:
            # no duplicate in sub-sampling
            choices = np.random.choice(
                point_idxs, self.num_points, replace=False)
        else:
            # do not use random choice here to avoid some points not counted
            dup = np.random.choice(point_idxs.size,
                                   self.num_points - point_idxs.size)
            idx_dup = np.concatenate(
                [np.arange(point_idxs.size),
                 np.array(dup)], 0)
            choices = point_idxs[idx_dup]
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323

        # construct model input
        points = self._input_generation(coords[choices], cur_center, coord_max,
                                        attributes[choices], attribute_dims,
                                        point_type)

        return points, choices

    def __call__(self, results):
        """Call function to sample points to in indoor scenes.

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
1324
            dict: Results after sampling, 'points', 'pts_instance_mask'
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
                and 'pts_semantic_mask' keys are updated in the result dict.
        """
        points = results['points']

        assert 'pts_semantic_mask' in results.keys(), \
            'semantic mask should be provided in training and evaluation'
        pts_semantic_mask = results['pts_semantic_mask']

        points, choices = self._patch_points_sampling(points,
                                                      pts_semantic_mask)

        results['points'] = points
        results['pts_semantic_mask'] = pts_semantic_mask[choices]
        pts_instance_mask = results.get('pts_instance_mask', None)
        if pts_instance_mask is not None:
            results['pts_instance_mask'] = pts_instance_mask[choices]

        return results

    def __repr__(self):
        """str: Return a string that describes the module."""
        repr_str = self.__class__.__name__
        repr_str += f'(num_points={self.num_points},'
        repr_str += f' block_size={self.block_size},'
        repr_str += f' ignore_index={self.ignore_index},'
        repr_str += f' use_normalized_coord={self.use_normalized_coord},'
1351
1352
        repr_str += f' num_try={self.num_try},'
        repr_str += f' enlarge_size={self.enlarge_size},'
1353
1354
        repr_str += f' min_unique_num={self.min_unique_num},'
        repr_str += f' eps={self.eps})'
wuyuefeng's avatar
wuyuefeng committed
1355
        return repr_str
1356
1357


1358
@TRANSFORMS.register_module()
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
class BackgroundPointsFilter(object):
    """Filter background points near the bounding box.

    Args:
        bbox_enlarge_range (tuple[float], float): Bbox enlarge range.
    """

    def __init__(self, bbox_enlarge_range):
        assert (is_tuple_of(bbox_enlarge_range, float)
                and len(bbox_enlarge_range) == 3) \
            or isinstance(bbox_enlarge_range, float), \
            f'Invalid arguments bbox_enlarge_range {bbox_enlarge_range}'

        if isinstance(bbox_enlarge_range, float):
            bbox_enlarge_range = [bbox_enlarge_range] * 3
        self.bbox_enlarge_range = np.array(
            bbox_enlarge_range, dtype=np.float32)[np.newaxis, :]

    def __call__(self, input_dict):
        """Call function to filter points by the range.

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
1384
            dict: Results after filtering, 'points', 'pts_instance_mask'
1385
                and 'pts_semantic_mask' keys are updated in the result dict.
1386
1387
1388
1389
        """
        points = input_dict['points']
        gt_bboxes_3d = input_dict['gt_bboxes_3d']

xiliu8006's avatar
xiliu8006 committed
1390
1391
1392
1393
        # avoid groundtruth being modified
        gt_bboxes_3d_np = gt_bboxes_3d.tensor.clone().numpy()
        gt_bboxes_3d_np[:, :3] = gt_bboxes_3d.gravity_center.clone().numpy()

1394
1395
        enlarged_gt_bboxes_3d = gt_bboxes_3d_np.copy()
        enlarged_gt_bboxes_3d[:, 3:6] += self.bbox_enlarge_range
xiliu8006's avatar
xiliu8006 committed
1396
        points_numpy = points.tensor.clone().numpy()
1397
1398
        foreground_masks = box_np_ops.points_in_rbbox(
            points_numpy, gt_bboxes_3d_np, origin=(0.5, 0.5, 0.5))
1399
        enlarge_foreground_masks = box_np_ops.points_in_rbbox(
1400
            points_numpy, enlarged_gt_bboxes_3d, origin=(0.5, 0.5, 0.5))
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
        foreground_masks = foreground_masks.max(1)
        enlarge_foreground_masks = enlarge_foreground_masks.max(1)
        valid_masks = ~np.logical_and(~foreground_masks,
                                      enlarge_foreground_masks)

        input_dict['points'] = points[valid_masks]
        pts_instance_mask = input_dict.get('pts_instance_mask', None)
        if pts_instance_mask is not None:
            input_dict['pts_instance_mask'] = pts_instance_mask[valid_masks]

        pts_semantic_mask = input_dict.get('pts_semantic_mask', None)
        if pts_semantic_mask is not None:
            input_dict['pts_semantic_mask'] = pts_semantic_mask[valid_masks]
        return input_dict

    def __repr__(self):
        """str: Return a string that describes the module."""
        repr_str = self.__class__.__name__
1419
        repr_str += f'(bbox_enlarge_range={self.bbox_enlarge_range.tolist()})'
1420
        return repr_str
1421
1422


1423
@TRANSFORMS.register_module()
1424
1425
1426
1427
1428
1429
1430
1431
class VoxelBasedPointSampler(object):
    """Voxel based point sampler.

    Apply voxel sampling to multiple sweep points.

    Args:
        cur_sweep_cfg (dict): Config for sampling current points.
        prev_sweep_cfg (dict): Config for sampling previous points.
1432
        time_dim (int): Index that indicate the time dimension
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
            for input points.
    """

    def __init__(self, cur_sweep_cfg, prev_sweep_cfg=None, time_dim=3):
        self.cur_voxel_generator = VoxelGenerator(**cur_sweep_cfg)
        self.cur_voxel_num = self.cur_voxel_generator._max_voxels
        self.time_dim = time_dim
        if prev_sweep_cfg is not None:
            assert prev_sweep_cfg['max_num_points'] == \
                cur_sweep_cfg['max_num_points']
            self.prev_voxel_generator = VoxelGenerator(**prev_sweep_cfg)
            self.prev_voxel_num = self.prev_voxel_generator._max_voxels
        else:
            self.prev_voxel_generator = None
            self.prev_voxel_num = 0

    def _sample_points(self, points, sampler, point_dim):
        """Sample points for each points subset.

        Args:
            points (np.ndarray): Points subset to be sampled.
            sampler (VoxelGenerator): Voxel based sampler for
                each points subset.
1456
            point_dim (int): The dimension of each points
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481

        Returns:
            np.ndarray: Sampled points.
        """
        voxels, coors, num_points_per_voxel = sampler.generate(points)
        if voxels.shape[0] < sampler._max_voxels:
            padding_points = np.zeros([
                sampler._max_voxels - voxels.shape[0], sampler._max_num_points,
                point_dim
            ],
                                      dtype=points.dtype)
            padding_points[:] = voxels[0]
            sample_points = np.concatenate([voxels, padding_points], axis=0)
        else:
            sample_points = voxels

        return sample_points

    def __call__(self, results):
        """Call function to sample points from multiple sweeps.

        Args:
            input_dict (dict): Result dict from loading pipeline.

        Returns:
1482
            dict: Results after sampling, 'points', 'pts_instance_mask'
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
                and 'pts_semantic_mask' keys are updated in the result dict.
        """
        points = results['points']
        original_dim = points.shape[1]

        # TODO: process instance and semantic mask while _max_num_points
        # is larger than 1
        # Extend points with seg and mask fields
        map_fields2dim = []
        start_dim = original_dim
1493
1494
        points_numpy = points.tensor.numpy()
        extra_channel = [points_numpy]
1495
1496
1497
1498
1499
1500
1501
1502
1503
        for idx, key in enumerate(results['pts_mask_fields']):
            map_fields2dim.append((key, idx + start_dim))
            extra_channel.append(results[key][..., None])

        start_dim += len(results['pts_mask_fields'])
        for idx, key in enumerate(results['pts_seg_fields']):
            map_fields2dim.append((key, idx + start_dim))
            extra_channel.append(results[key][..., None])

1504
        points_numpy = np.concatenate(extra_channel, axis=-1)
1505
1506
1507
1508
1509

        # Split points into two part, current sweep points and
        # previous sweeps points.
        # TODO: support different sampling methods for next sweeps points
        # and previous sweeps points.
1510
1511
1512
        cur_points_flag = (points_numpy[:, self.time_dim] == 0)
        cur_sweep_points = points_numpy[cur_points_flag]
        prev_sweeps_points = points_numpy[~cur_points_flag]
1513
1514
1515
1516
1517
1518
1519
1520
1521
        if prev_sweeps_points.shape[0] == 0:
            prev_sweeps_points = cur_sweep_points

        # Shuffle points before sampling
        np.random.shuffle(cur_sweep_points)
        np.random.shuffle(prev_sweeps_points)

        cur_sweep_points = self._sample_points(cur_sweep_points,
                                               self.cur_voxel_generator,
1522
                                               points_numpy.shape[1])
1523
1524
1525
        if self.prev_voxel_generator is not None:
            prev_sweeps_points = self._sample_points(prev_sweeps_points,
                                                     self.prev_voxel_generator,
1526
                                                     points_numpy.shape[1])
1527

1528
1529
            points_numpy = np.concatenate(
                [cur_sweep_points, prev_sweeps_points], 0)
1530
        else:
1531
            points_numpy = cur_sweep_points
1532
1533

        if self.cur_voxel_generator._max_num_points == 1:
1534
1535
            points_numpy = points_numpy.squeeze(1)
        results['points'] = points.new_point(points_numpy[..., :original_dim])
1536

1537
        # Restore the corresponding seg and mask fields
1538
        for key, dim_index in map_fields2dim:
1539
            results[key] = points_numpy[..., dim_index]
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562

        return results

    def __repr__(self):
        """str: Return a string that describes the module."""

        def _auto_indent(repr_str, indent):
            repr_str = repr_str.split('\n')
            repr_str = [' ' * indent + t + '\n' for t in repr_str]
            repr_str = ''.join(repr_str)[:-1]
            return repr_str

        repr_str = self.__class__.__name__
        indent = 4
        repr_str += '(\n'
        repr_str += ' ' * indent + f'num_cur_sweep={self.cur_voxel_num},\n'
        repr_str += ' ' * indent + f'num_prev_sweep={self.prev_voxel_num},\n'
        repr_str += ' ' * indent + f'time_dim={self.time_dim},\n'
        repr_str += ' ' * indent + 'cur_voxel_generator=\n'
        repr_str += f'{_auto_indent(repr(self.cur_voxel_generator), 8)},\n'
        repr_str += ' ' * indent + 'prev_voxel_generator=\n'
        repr_str += f'{_auto_indent(repr(self.prev_voxel_generator), 8)})'
        return repr_str
1563
1564


1565
@TRANSFORMS.register_module()
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
class AffineResize(object):
    """Get the affine transform matrices to the target size.

    Different from :class:`RandomAffine` in MMDetection, this class can
    calculate the affine transform matrices while resizing the input image
    to a fixed size. The affine transform matrices include: 1) matrix
    transforming original image to the network input image size. 2) matrix
    transforming original image to the network output feature map size.

    Args:
        img_scale (tuple): Images scales for resizing.
        down_ratio (int): The down ratio of feature map.
            Actually the arg should be >= 1.
        bbox_clip_border (bool, optional): Whether clip the objects
            outside the border of the image. Defaults to True.
    """

    def __init__(self, img_scale, down_ratio, bbox_clip_border=True):

        self.img_scale = img_scale
        self.down_ratio = down_ratio
        self.bbox_clip_border = bbox_clip_border

    def __call__(self, results):
        """Call function to do affine transform to input image and labels.

        Args:
            results (dict): Result dict from loading pipeline.

        Returns:
            dict: Results after affine resize, 'affine_aug', 'trans_mat'
                keys are added in the result dict.
        """
        # The results have gone through RandomShiftScale before AffineResize
        if 'center' not in results:
            img = results['img']
            height, width = img.shape[:2]
            center = np.array([width / 2, height / 2], dtype=np.float32)
            size = np.array([width, height], dtype=np.float32)
            results['affine_aug'] = False
        else:
            # The results did not go through RandomShiftScale before
            # AffineResize
            img = results['img']
            center = results['center']
            size = results['size']

        trans_affine = self._get_transform_matrix(center, size, self.img_scale)

        img = cv2.warpAffine(img, trans_affine[:2, :], self.img_scale)

        if isinstance(self.down_ratio, tuple):
            trans_mat = [
                self._get_transform_matrix(
                    center, size,
                    (self.img_scale[0] // ratio, self.img_scale[1] // ratio))
                for ratio in self.down_ratio
            ]  # (3, 3)
        else:
            trans_mat = self._get_transform_matrix(
                center, size, (self.img_scale[0] // self.down_ratio,
                               self.img_scale[1] // self.down_ratio))

        results['img'] = img
        results['img_shape'] = img.shape
        results['pad_shape'] = img.shape
        results['trans_mat'] = trans_mat

        self._affine_bboxes(results, trans_affine)

        if 'centers2d' in results:
            centers2d = self._affine_transform(results['centers2d'],
                                               trans_affine)
            valid_index = (centers2d[:, 0] >
                           0) & (centers2d[:, 0] <
                                 self.img_scale[0]) & (centers2d[:, 1] > 0) & (
                                     centers2d[:, 1] < self.img_scale[1])
            results['centers2d'] = centers2d[valid_index]

            for key in results.get('bbox_fields', []):
                if key in ['gt_bboxes']:
                    results[key] = results[key][valid_index]
                    if 'gt_labels' in results:
                        results['gt_labels'] = results['gt_labels'][
                            valid_index]
                    if 'gt_masks' in results:
                        raise NotImplementedError(
                            'AffineResize only supports bbox.')

            for key in results.get('bbox3d_fields', []):
                if key in ['gt_bboxes_3d']:
                    results[key].tensor = results[key].tensor[valid_index]
                    if 'gt_labels_3d' in results:
                        results['gt_labels_3d'] = results['gt_labels_3d'][
                            valid_index]

            results['depths'] = results['depths'][valid_index]

        return results

    def _affine_bboxes(self, results, matrix):
        """Affine transform bboxes to input image.

        Args:
            results (dict): Result dict from loading pipeline.
            matrix (np.ndarray): Matrix transforming original
                image to the network input image size.
                shape: (3, 3)
        """

        for key in results.get('bbox_fields', []):
            bboxes = results[key]
            bboxes[:, :2] = self._affine_transform(bboxes[:, :2], matrix)
            bboxes[:, 2:] = self._affine_transform(bboxes[:, 2:], matrix)
            if self.bbox_clip_border:
                bboxes[:,
                       [0, 2]] = bboxes[:,
                                        [0, 2]].clip(0, self.img_scale[0] - 1)
                bboxes[:,
                       [1, 3]] = bboxes[:,
                                        [1, 3]].clip(0, self.img_scale[1] - 1)
            results[key] = bboxes

    def _affine_transform(self, points, matrix):
1690
        """Affine transform bbox points to input image.
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743

        Args:
            points (np.ndarray): Points to be transformed.
                shape: (N, 2)
            matrix (np.ndarray): Affine transform matrix.
                shape: (3, 3)

        Returns:
            np.ndarray: Transformed points.
        """
        num_points = points.shape[0]
        hom_points_2d = np.concatenate((points, np.ones((num_points, 1))),
                                       axis=1)
        hom_points_2d = hom_points_2d.T
        affined_points = np.matmul(matrix, hom_points_2d).T
        return affined_points[:, :2]

    def _get_transform_matrix(self, center, scale, output_scale):
        """Get affine transform matrix.

        Args:
            center (tuple): Center of current image.
            scale (tuple): Scale of current image.
            output_scale (tuple[float]): The transform target image scales.

        Returns:
            np.ndarray: Affine transform matrix.
        """
        # TODO: further add rot and shift here.
        src_w = scale[0]
        dst_w = output_scale[0]
        dst_h = output_scale[1]

        src_dir = np.array([0, src_w * -0.5])
        dst_dir = np.array([0, dst_w * -0.5])

        src = np.zeros((3, 2), dtype=np.float32)
        dst = np.zeros((3, 2), dtype=np.float32)
        src[0, :] = center
        src[1, :] = center + src_dir
        dst[0, :] = np.array([dst_w * 0.5, dst_h * 0.5])
        dst[1, :] = np.array([dst_w * 0.5, dst_h * 0.5]) + dst_dir

        src[2, :] = self._get_ref_point(src[0, :], src[1, :])
        dst[2, :] = self._get_ref_point(dst[0, :], dst[1, :])

        get_matrix = cv2.getAffineTransform(src, dst)

        matrix = np.concatenate((get_matrix, [[0., 0., 1.]]))

        return matrix.astype(np.float32)

    def _get_ref_point(self, ref_point1, ref_point2):
1744
        """Get reference point to calculate affine transform matrix.
1745
1746

        While using opencv to calculate the affine matrix, we need at least
1747
        three corresponding points separately on original image and target
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
        image. Here we use two points to get the the third reference point.
        """
        d = ref_point1 - ref_point2
        ref_point3 = ref_point2 + np.array([-d[1], d[0]])
        return ref_point3

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


1761
@TRANSFORMS.register_module()
1762
1763
1764
1765
1766
class RandomShiftScale(object):
    """Random shift scale.

    Different from the normal shift and scale function, it doesn't
    directly shift or scale image. It can record the shift and scale
1767
    infos into loading TRANSFORMS. It's designed to be used with
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
    AffineResize together.

    Args:
        shift_scale (tuple[float]): Shift and scale range.
        aug_prob (float): The shifting and scaling probability.
    """

    def __init__(self, shift_scale, aug_prob):

        self.shift_scale = shift_scale
        self.aug_prob = aug_prob

    def __call__(self, results):
        """Call function to record random shift and scale infos.

        Args:
            results (dict): Result dict from loading pipeline.

        Returns:
            dict: Results after random shift and scale, 'center', 'size'
                and 'affine_aug' keys are added in the result dict.
        """
        img = results['img']

        height, width = img.shape[:2]

        center = np.array([width / 2, height / 2], dtype=np.float32)
        size = np.array([width, height], dtype=np.float32)

        if random.random() < self.aug_prob:
            shift, scale = self.shift_scale[0], self.shift_scale[1]
            shift_ranges = np.arange(-shift, shift + 0.1, 0.1)
            center[0] += size[0] * random.choice(shift_ranges)
            center[1] += size[1] * random.choice(shift_ranges)
            scale_ranges = np.arange(1 - scale, 1 + scale + 0.1, 0.1)
            size *= random.choice(scale_ranges)
            results['affine_aug'] = True
        else:
            results['affine_aug'] = False

        results['center'] = center
        results['size'] = size

        return results

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