transforms_3d.py 55 KB
Newer Older
dingchang's avatar
dingchang committed
1
# Copyright (c) OpenMMLab. All rights reserved.
zhangwenwei's avatar
zhangwenwei committed
2
import numpy as np
3
import warnings
4
from mmcv import is_tuple_of
5
from mmcv.utils import build_from_cfg
zhangwenwei's avatar
zhangwenwei committed
6

7
from mmdet3d.core import VoxelGenerator
8
9
from mmdet3d.core.bbox import (CameraInstance3DBoxes, DepthInstance3DBoxes,
                               LiDARInstance3DBoxes, box_np_ops)
10
from mmdet.datasets.builder import PIPELINES
zhangwenwei's avatar
zhangwenwei committed
11
from mmdet.datasets.pipelines import RandomFlip
12
from ..builder import OBJECTSAMPLERS
zhangwenwei's avatar
zhangwenwei committed
13
14
15
from .data_augment_utils import noise_per_object_v3_


16
17
18
19
20
21
22
23
24
@PIPELINES.register_module()
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:
25
        drop_ratio (float, optional): The probability of dropping point colors.
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
            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:
41
            dict: Results after color dropping,
42
43
44
45
46
47
48
                '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'

49
50
51
52
53
54
55
        # 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:
56
57
58
59
60
61
62
63
64
65
            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


66
@PIPELINES.register_module()
zhangwenwei's avatar
zhangwenwei committed
67
68
69
70
71
72
73
74
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.

    Args:
zhangwenwei's avatar
zhangwenwei committed
75
76
77
        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
78
            to that of 2D images. Defaults to True.
wuyuefeng's avatar
wuyuefeng committed
79
        flip_ratio_bev_horizontal (float, optional): The flipping probability
liyinhao's avatar
liyinhao committed
80
            in horizontal direction. Defaults to 0.0.
wuyuefeng's avatar
wuyuefeng committed
81
        flip_ratio_bev_vertical (float, optional): The flipping probability
liyinhao's avatar
liyinhao committed
82
            in vertical direction. Defaults to 0.0.
zhangwenwei's avatar
zhangwenwei committed
83
84
    """

wuyuefeng's avatar
wuyuefeng committed
85
86
87
88
89
90
91
    def __init__(self,
                 sync_2d=True,
                 flip_ratio_bev_horizontal=0.0,
                 flip_ratio_bev_vertical=0.0,
                 **kwargs):
        super(RandomFlip3D, self).__init__(
            flip_ratio=flip_ratio_bev_horizontal, **kwargs)
zhangwenwei's avatar
zhangwenwei committed
92
        self.sync_2d = sync_2d
wuyuefeng's avatar
wuyuefeng committed
93
94
95
96
97
98
99
100
101
102
103
        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

    def random_flip_data_3d(self, input_dict, direction='horizontal'):
104
105
106
107
        """Flip 3D data randomly.

        Args:
            input_dict (dict): Result dict from loading pipeline.
108
109
            direction (str, optional): Flip direction.
                Default: 'horizontal'.
110
111

        Returns:
112
            dict: Flipped results, 'points', 'bbox3d_fields' keys are
113
114
                updated in the result dict.
        """
wuyuefeng's avatar
wuyuefeng committed
115
        assert direction in ['horizontal', 'vertical']
116
117
118
119
120
        if len(input_dict['bbox3d_fields']) == 0:  # test mode
            input_dict['bbox3d_fields'].append('empty_box3d')
            input_dict['empty_box3d'] = input_dict['box_type_3d'](
                np.array([], dtype=np.float32))
        assert len(input_dict['bbox3d_fields']) == 1
zhangwenwei's avatar
zhangwenwei committed
121
        for key in input_dict['bbox3d_fields']:
122
123
124
125
126
127
128
129
            if 'points' in input_dict:
                input_dict['points'] = input_dict[key].flip(
                    direction, points=input_dict['points'])
            else:
                input_dict[key].flip(direction)
        if 'centers2d' in input_dict:
            assert self.sync_2d is True and direction == 'horizontal', \
                'Only support sync_2d=True and horizontal flip with images'
130
            w = input_dict['ori_shape'][1]
131
132
            input_dict['centers2d'][..., 0] = \
                w - input_dict['centers2d'][..., 0]
133
134
            # need to modify the horizontal position of camera center
            # along u-axis in the image (flip like centers2d)
135
            # ['cam2img'][0][2] = c_u
136
137
            # see more details and examples at
            # https://github.com/open-mmlab/mmdetection3d/pull/744
138
            input_dict['cam2img'][0][2] = w - input_dict['cam2img'][0][2]
zhangwenwei's avatar
zhangwenwei committed
139
140

    def __call__(self, input_dict):
141
        """Call function to flip points, values in the ``bbox3d_fields`` and
142
143
144
145
146
147
        also flip 2D image and its annotations.

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

        Returns:
148
149
            dict: Flipped results, 'flip', 'flip_direction',
                'pcd_horizontal_flip' and 'pcd_vertical_flip' keys are added
150
151
                into result dict.
        """
zhangwenwei's avatar
zhangwenwei committed
152
        # filp 2D image and its annotations
zhangwenwei's avatar
zhangwenwei committed
153
        super(RandomFlip3D, self).__call__(input_dict)
zhangwenwei's avatar
zhangwenwei committed
154

zhangwenwei's avatar
zhangwenwei committed
155
        if self.sync_2d:
wuyuefeng's avatar
wuyuefeng committed
156
157
            input_dict['pcd_horizontal_flip'] = input_dict['flip']
            input_dict['pcd_vertical_flip'] = False
zhangwenwei's avatar
zhangwenwei committed
158
        else:
wuyuefeng's avatar
wuyuefeng committed
159
160
161
162
163
164
165
166
167
            if 'pcd_horizontal_flip' not in input_dict:
                flip_horizontal = True if np.random.rand(
                ) < self.flip_ratio else False
                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

168
169
170
        if 'transformation_3d_flow' not in input_dict:
            input_dict['transformation_3d_flow'] = []

wuyuefeng's avatar
wuyuefeng committed
171
172
        if input_dict['pcd_horizontal_flip']:
            self.random_flip_data_3d(input_dict, 'horizontal')
173
            input_dict['transformation_3d_flow'].extend(['HF'])
wuyuefeng's avatar
wuyuefeng committed
174
175
        if input_dict['pcd_vertical_flip']:
            self.random_flip_data_3d(input_dict, 'vertical')
176
            input_dict['transformation_3d_flow'].extend(['VF'])
zhangwenwei's avatar
zhangwenwei committed
177
178
        return input_dict

zhangwenwei's avatar
zhangwenwei committed
179
    def __repr__(self):
180
        """str: Return a string that describes the module."""
wuyuefeng's avatar
wuyuefeng committed
181
        repr_str = self.__class__.__name__
182
        repr_str += f'(sync_2d={self.sync_2d},'
183
        repr_str += f' flip_ratio_bev_vertical={self.flip_ratio_bev_vertical})'
wuyuefeng's avatar
wuyuefeng committed
184
        return repr_str
zhangwenwei's avatar
zhangwenwei committed
185

zhangwenwei's avatar
zhangwenwei committed
186

187
188
189
190
@PIPELINES.register_module()
class RandomJitterPoints(object):
    """Randomly jitter point coordinates.

191
    Different from the global translation in ``GlobalRotScaleTrans``, here we
192
193
194
195
        apply different noises to each point in a scene.

    Args:
        jitter_std (list[float]): The standard deviation of jittering noise.
196
197
            This applies random noise to all points in a 3D scene, which is
            sampled from a gaussian distribution whose standard deviation is
198
            set by ``jitter_std``. Defaults to [0.01, 0.01, 0.01]
199
        clip_range (list[float]): Clip the randomly generated jitter
200
201
202
203
            noise into this range. If None is given, don't perform clipping.
            Defaults to [-0.05, 0.05]

    Note:
204
        This transform should only be used in point cloud segmentation tasks
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
            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:
233
            dict: Results after adding noise to each point,
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
                '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


255
@PIPELINES.register_module()
zhangwenwei's avatar
zhangwenwei committed
256
class ObjectSample(object):
zhangwenwei's avatar
zhangwenwei committed
257
    """Sample GT objects to the data.
zhangwenwei's avatar
zhangwenwei committed
258
259
260
261
262

    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
263
            Defaults to False.
zhangwenwei's avatar
zhangwenwei committed
264
    """
zhangwenwei's avatar
zhangwenwei committed
265
266
267
268
269
270
271
272
273
274

    def __init__(self, db_sampler, sample_2d=False):
        self.sampler_cfg = db_sampler
        self.sample_2d = sample_2d
        if 'type' not in db_sampler.keys():
            db_sampler['type'] = 'DataBaseSampler'
        self.db_sampler = build_from_cfg(db_sampler, OBJECTSAMPLERS)

    @staticmethod
    def remove_points_in_boxes(points, boxes):
275
276
277
        """Remove the points in the sampled bounding boxes.

        Args:
278
            points (:obj:`BasePoints`): Input point cloud array.
279
280
281
282
283
            boxes (np.ndarray): Sampled ground truth boxes.

        Returns:
            np.ndarray: Points with those in the boxes removed.
        """
284
        masks = box_np_ops.points_in_rbbox(points.coord.numpy(), boxes)
zhangwenwei's avatar
zhangwenwei committed
285
286
287
288
        points = points[np.logical_not(masks.any(-1))]
        return points

    def __call__(self, input_dict):
289
290
291
292
293
294
        """Call function to sample ground truth objects to the data.

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

        Returns:
295
296
            dict: Results after object sampling augmentation,
                'points', 'gt_bboxes_3d', 'gt_labels_3d' keys are updated
297
298
                in the result dict.
        """
zhangwenwei's avatar
zhangwenwei committed
299
        gt_bboxes_3d = input_dict['gt_bboxes_3d']
zhangwenwei's avatar
zhangwenwei committed
300
301
        gt_labels_3d = input_dict['gt_labels_3d']

zhangwenwei's avatar
zhangwenwei committed
302
303
304
        # change to float for blending operation
        points = input_dict['points']
        if self.sample_2d:
wuyuefeng's avatar
wuyuefeng committed
305
            img = input_dict['img']
zhangwenwei's avatar
zhangwenwei committed
306
307
308
            gt_bboxes_2d = input_dict['gt_bboxes']
            # Assume for now 3D & 2D bboxes are the same
            sampled_dict = self.db_sampler.sample_all(
309
310
311
312
                gt_bboxes_3d.tensor.numpy(),
                gt_labels_3d,
                gt_bboxes_2d=gt_bboxes_2d,
                img=img)
zhangwenwei's avatar
zhangwenwei committed
313
314
        else:
            sampled_dict = self.db_sampler.sample_all(
315
                gt_bboxes_3d.tensor.numpy(), gt_labels_3d, img=None)
zhangwenwei's avatar
zhangwenwei committed
316
317
318
319

        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
320
            sampled_gt_labels = sampled_dict['gt_labels_3d']
zhangwenwei's avatar
zhangwenwei committed
321

zhangwenwei's avatar
zhangwenwei committed
322
323
            gt_labels_3d = np.concatenate([gt_labels_3d, sampled_gt_labels],
                                          axis=0)
324
325
326
            gt_bboxes_3d = gt_bboxes_3d.new_box(
                np.concatenate(
                    [gt_bboxes_3d.tensor.numpy(), sampled_gt_bboxes_3d]))
zhangwenwei's avatar
zhangwenwei committed
327

zhangwenwei's avatar
zhangwenwei committed
328
329
            points = self.remove_points_in_boxes(points, sampled_gt_bboxes_3d)
            # check the points dimension
330
            points = points.cat([sampled_points, points])
zhangwenwei's avatar
zhangwenwei committed
331
332
333
334
335

            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
336

zhangwenwei's avatar
zhangwenwei committed
337
                input_dict['gt_bboxes'] = gt_bboxes_2d
wuyuefeng's avatar
wuyuefeng committed
338
                input_dict['img'] = sampled_dict['img']
zhangwenwei's avatar
zhangwenwei committed
339
340

        input_dict['gt_bboxes_3d'] = gt_bboxes_3d
341
        input_dict['gt_labels_3d'] = gt_labels_3d.astype(np.long)
zhangwenwei's avatar
zhangwenwei committed
342
        input_dict['points'] = points
zhangwenwei's avatar
zhangwenwei committed
343

zhangwenwei's avatar
zhangwenwei committed
344
345
346
        return input_dict

    def __repr__(self):
347
        """str: Return a string that describes the module."""
348
349
350
351
352
353
354
355
356
        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
357
358


359
@PIPELINES.register_module()
zhangwenwei's avatar
zhangwenwei committed
360
class ObjectNoise(object):
zhangwenwei's avatar
zhangwenwei committed
361
    """Apply noise to each GT objects in the scene.
zhangwenwei's avatar
zhangwenwei committed
362
363

    Args:
364
        translation_std (list[float], optional): Standard deviation of the
zhangwenwei's avatar
zhangwenwei committed
365
366
            distribution where translation noise are sampled from.
            Defaults to [0.25, 0.25, 0.25].
367
        global_rot_range (list[float], optional): Global rotation to the scene.
zhangwenwei's avatar
zhangwenwei committed
368
            Defaults to [0.0, 0.0].
369
        rot_range (list[float], optional): Object rotation range.
zhangwenwei's avatar
zhangwenwei committed
370
371
372
373
            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
374
375

    def __init__(self,
zhangwenwei's avatar
zhangwenwei committed
376
                 translation_std=[0.25, 0.25, 0.25],
zhangwenwei's avatar
zhangwenwei committed
377
                 global_rot_range=[0.0, 0.0],
zhangwenwei's avatar
zhangwenwei committed
378
                 rot_range=[-0.15707963267, 0.15707963267],
zhangwenwei's avatar
zhangwenwei committed
379
                 num_try=100):
zhangwenwei's avatar
zhangwenwei committed
380
        self.translation_std = translation_std
zhangwenwei's avatar
zhangwenwei committed
381
        self.global_rot_range = global_rot_range
zhangwenwei's avatar
zhangwenwei committed
382
        self.rot_range = rot_range
zhangwenwei's avatar
zhangwenwei committed
383
384
385
        self.num_try = num_try

    def __call__(self, input_dict):
386
387
388
389
390
391
        """Call function to apply noise to each ground truth in the scene.

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

        Returns:
392
            dict: Results after adding noise to each object,
393
394
                'points', 'gt_bboxes_3d' keys are updated in the result dict.
        """
zhangwenwei's avatar
zhangwenwei committed
395
396
        gt_bboxes_3d = input_dict['gt_bboxes_3d']
        points = input_dict['points']
zhangwenwei's avatar
zhangwenwei committed
397

398
        # TODO: this is inplace operation
399
        numpy_box = gt_bboxes_3d.tensor.numpy()
400
401
        numpy_points = points.tensor.numpy()

zhangwenwei's avatar
zhangwenwei committed
402
        noise_per_object_v3_(
403
            numpy_box,
404
            numpy_points,
zhangwenwei's avatar
zhangwenwei committed
405
406
            rotation_perturb=self.rot_range,
            center_noise_std=self.translation_std,
zhangwenwei's avatar
zhangwenwei committed
407
408
            global_random_rot_range=self.global_rot_range,
            num_try=self.num_try)
409
410

        input_dict['gt_bboxes_3d'] = gt_bboxes_3d.new_box(numpy_box)
411
        input_dict['points'] = points.new_point(numpy_points)
zhangwenwei's avatar
zhangwenwei committed
412
413
414
        return input_dict

    def __repr__(self):
415
        """str: Return a string that describes the module."""
zhangwenwei's avatar
zhangwenwei committed
416
        repr_str = self.__class__.__name__
417
418
419
420
        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
421
422
423
        return repr_str


424
425
426
427
428
429
430
431
@PIPELINES.register_module()
class GlobalAlignment(object):
    """Apply global alignment to 3D scene points by rotation and translation.

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

    Note:
432
433
        We do not record the applied rotation and translation as in
            GlobalRotScaleTrans. Because usually, we do not need to reverse
434
            the alignment step.
435
        For example, ScanNet 3D detection task uses aligned ground-truth
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
            bounding boxes for evaluation.
    """

    def __init__(self, rotation_axis):
        self.rotation_axis = rotation_axis

    def _trans_points(self, input_dict, trans_factor):
        """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.
        """
        input_dict['points'].translate(trans_factor)

    def _rot_points(self, input_dict, rot_mat):
        """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
        input_dict['points'].rotate(rot_mat.T)

    def _check_rot_mat(self, rot_mat):
        """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}'

    def __call__(self, input_dict):
        """Call function to shuffle points.

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

        Returns:
487
            dict: Results after global alignment, 'points' and keys in
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
                input_dict['bbox3d_fields'] are updated in the result dict.
        """
        assert 'axis_align_matrix' in input_dict['ann_info'].keys(), \
            'axis_align_matrix is not provided in GlobalAlignment'

        axis_align_matrix = input_dict['ann_info']['axis_align_matrix']
        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)
        self._rot_points(input_dict, rot_mat)
        self._trans_points(input_dict, trans_vec)

        return input_dict

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


511
@PIPELINES.register_module()
zhangwenwei's avatar
zhangwenwei committed
512
class GlobalRotScaleTrans(object):
zhangwenwei's avatar
zhangwenwei committed
513
    """Apply global rotation, scaling and translation to a 3D scene.
zhangwenwei's avatar
zhangwenwei committed
514
515

    Args:
516
        rot_range (list[float], optional): Range of rotation angle.
liyinhao's avatar
liyinhao committed
517
            Defaults to [-0.78539816, 0.78539816] (close to [-pi/4, pi/4]).
518
        scale_ratio_range (list[float], optional): Range of scale ratio.
liyinhao's avatar
liyinhao committed
519
            Defaults to [0.95, 1.05].
520
521
        translation_std (list[float], optional): The standard deviation of
            translation noise applied to a scene, which
zhangwenwei's avatar
zhangwenwei committed
522
            is sampled from a gaussian distribution whose standard deviation
liyinhao's avatar
liyinhao committed
523
            is set by ``translation_std``. Defaults to [0, 0, 0]
524
        shift_height (bool, optional): Whether to shift height.
wuyuefeng's avatar
wuyuefeng committed
525
            (the fourth dimension of indoor points) when scaling.
liyinhao's avatar
liyinhao committed
526
            Defaults to False.
zhangwenwei's avatar
zhangwenwei committed
527
    """
zhangwenwei's avatar
zhangwenwei committed
528
529

    def __init__(self,
zhangwenwei's avatar
zhangwenwei committed
530
531
                 rot_range=[-0.78539816, 0.78539816],
                 scale_ratio_range=[0.95, 1.05],
wuyuefeng's avatar
wuyuefeng committed
532
533
                 translation_std=[0, 0, 0],
                 shift_height=False):
534
535
536
537
538
        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
539
        self.rot_range = rot_range
540
541
542

        assert isinstance(scale_ratio_range, seq_types), \
            f'unsupported scale_ratio_range type {type(scale_ratio_range)}'
zhangwenwei's avatar
zhangwenwei committed
543
        self.scale_ratio_range = scale_ratio_range
544
545
546
547
548
549
550

        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
            ]
551
552
        assert all([std >= 0 for std in translation_std]), \
            'translation_std should be positive'
zhangwenwei's avatar
zhangwenwei committed
553
        self.translation_std = translation_std
wuyuefeng's avatar
wuyuefeng committed
554
        self.shift_height = shift_height
zhangwenwei's avatar
zhangwenwei committed
555
556

    def _trans_bbox_points(self, input_dict):
557
558
559
560
561
562
        """Private function to translate bounding boxes and points.

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

        Returns:
563
564
            dict: Results after translation, 'points', 'pcd_trans'
                and keys in input_dict['bbox3d_fields'] are updated
565
566
                in the result dict.
        """
567
        translation_std = np.array(self.translation_std, dtype=np.float32)
zhangwenwei's avatar
zhangwenwei committed
568
569
        trans_factor = np.random.normal(scale=translation_std, size=3).T

570
        input_dict['points'].translate(trans_factor)
zhangwenwei's avatar
zhangwenwei committed
571
572
573
574
575
        input_dict['pcd_trans'] = trans_factor
        for key in input_dict['bbox3d_fields']:
            input_dict[key].translate(trans_factor)

    def _rot_bbox_points(self, input_dict):
576
577
578
579
580
581
        """Private function to rotate bounding boxes and points.

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

        Returns:
582
583
            dict: Results after rotation, 'points', 'pcd_rotation'
                and keys in input_dict['bbox3d_fields'] are updated
584
585
                in the result dict.
        """
zhangwenwei's avatar
zhangwenwei committed
586
        rotation = self.rot_range
zhangwenwei's avatar
zhangwenwei committed
587
        noise_rotation = np.random.uniform(rotation[0], rotation[1])
zhangwenwei's avatar
zhangwenwei committed
588

589
590
591
592
        # if no bbox in input_dict, only rotate points
        if len(input_dict['bbox3d_fields']) == 0:
            rot_mat_T = input_dict['points'].rotate(noise_rotation)
            input_dict['pcd_rotation'] = rot_mat_T
593
            input_dict['pcd_rotation_angle'] = noise_rotation
594
595
596
            return

        # rotate points with bboxes
zhangwenwei's avatar
zhangwenwei committed
597
        for key in input_dict['bbox3d_fields']:
wuyuefeng's avatar
wuyuefeng committed
598
599
600
601
602
            if len(input_dict[key].tensor) != 0:
                points, rot_mat_T = input_dict[key].rotate(
                    noise_rotation, input_dict['points'])
                input_dict['points'] = points
                input_dict['pcd_rotation'] = rot_mat_T
603
                input_dict['pcd_rotation_angle'] = noise_rotation
604

zhangwenwei's avatar
zhangwenwei committed
605
    def _scale_bbox_points(self, input_dict):
606
607
608
609
610
611
        """Private function to scale bounding boxes and points.

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

        Returns:
612
            dict: Results after scaling, 'points'and keys in
613
614
                input_dict['bbox3d_fields'] are updated in the result dict.
        """
zhangwenwei's avatar
zhangwenwei committed
615
        scale = input_dict['pcd_scale_factor']
616
617
        points = input_dict['points']
        points.scale(scale)
wuyuefeng's avatar
wuyuefeng committed
618
        if self.shift_height:
619
620
            assert 'height' in points.attribute_dims.keys(), \
                'setting shift_height=True but points have no height attribute'
621
622
            points.tensor[:, points.attribute_dims['height']] *= scale
        input_dict['points'] = points
wuyuefeng's avatar
wuyuefeng committed
623

zhangwenwei's avatar
zhangwenwei committed
624
625
        for key in input_dict['bbox3d_fields']:
            input_dict[key].scale(scale)
zhangwenwei's avatar
zhangwenwei committed
626

zhangwenwei's avatar
zhangwenwei committed
627
    def _random_scale(self, input_dict):
628
629
630
631
632
633
        """Private function to randomly set the scale factor.

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

        Returns:
634
            dict: Results after scaling, 'pcd_scale_factor' are updated
635
636
                in the result dict.
        """
zhangwenwei's avatar
zhangwenwei committed
637
638
639
        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
640
641

    def __call__(self, input_dict):
642
        """Private function to rotate, scale and translate bounding boxes and
643
644
645
646
647
648
649
        points.

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

        Returns:
            dict: Results after scaling, 'points', 'pcd_rotation',
650
                'pcd_scale_factor', 'pcd_trans' and keys in
651
652
                input_dict['bbox3d_fields'] are updated in the result dict.
        """
653
654
655
        if 'transformation_3d_flow' not in input_dict:
            input_dict['transformation_3d_flow'] = []

zhangwenwei's avatar
zhangwenwei committed
656
        self._rot_bbox_points(input_dict)
zhangwenwei's avatar
zhangwenwei committed
657

zhangwenwei's avatar
zhangwenwei committed
658
659
660
        if 'pcd_scale_factor' not in input_dict:
            self._random_scale(input_dict)
        self._scale_bbox_points(input_dict)
zhangwenwei's avatar
zhangwenwei committed
661

zhangwenwei's avatar
zhangwenwei committed
662
        self._trans_bbox_points(input_dict)
663
664

        input_dict['transformation_3d_flow'].extend(['R', 'S', 'T'])
zhangwenwei's avatar
zhangwenwei committed
665
666
667
        return input_dict

    def __repr__(self):
668
        """str: Return a string that describes the module."""
zhangwenwei's avatar
zhangwenwei committed
669
        repr_str = self.__class__.__name__
670
671
672
673
        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
674
675
676
        return repr_str


677
@PIPELINES.register_module()
zhangwenwei's avatar
zhangwenwei committed
678
class PointShuffle(object):
679
    """Shuffle input points."""
zhangwenwei's avatar
zhangwenwei committed
680
681

    def __call__(self, input_dict):
682
683
684
685
686
687
        """Call function to shuffle points.

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

        Returns:
688
            dict: Results after filtering, 'points', 'pts_instance_mask'
689
                and 'pts_semantic_mask' keys are updated in the result dict.
690
        """
691
692
693
694
695
696
697
698
699
700
701
702
        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
703
704
705
706
707
708
        return input_dict

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


709
@PIPELINES.register_module()
zhangwenwei's avatar
zhangwenwei committed
710
class ObjectRangeFilter(object):
711
712
713
714
715
    """Filter objects by the range.

    Args:
        point_cloud_range (list[float]): Point cloud range.
    """
zhangwenwei's avatar
zhangwenwei committed
716
717
718
719
720

    def __init__(self, point_cloud_range):
        self.pcd_range = np.array(point_cloud_range, dtype=np.float32)

    def __call__(self, input_dict):
721
722
723
724
725
726
        """Call function to filter objects by the range.

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

        Returns:
727
            dict: Results after filtering, 'gt_bboxes_3d', 'gt_labels_3d'
728
729
                keys are updated in the result dict.
        """
730
731
732
733
734
735
736
        # 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
737
        gt_bboxes_3d = input_dict['gt_bboxes_3d']
zhangwenwei's avatar
zhangwenwei committed
738
        gt_labels_3d = input_dict['gt_labels_3d']
739
        mask = gt_bboxes_3d.in_range_bev(bev_range)
zhangwenwei's avatar
zhangwenwei committed
740
        gt_bboxes_3d = gt_bboxes_3d[mask]
ZwwWayne's avatar
ZwwWayne committed
741
742
743
744
745
        # 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
746
747

        # limit rad to [-pi, pi]
748
749
        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
750
751
        input_dict['gt_labels_3d'] = gt_labels_3d

zhangwenwei's avatar
zhangwenwei committed
752
753
754
        return input_dict

    def __repr__(self):
755
        """str: Return a string that describes the module."""
zhangwenwei's avatar
zhangwenwei committed
756
        repr_str = self.__class__.__name__
757
        repr_str += f'(point_cloud_range={self.pcd_range.tolist()})'
zhangwenwei's avatar
zhangwenwei committed
758
759
760
        return repr_str


761
@PIPELINES.register_module()
zhangwenwei's avatar
zhangwenwei committed
762
class PointsRangeFilter(object):
763
764
765
766
767
    """Filter points by the range.

    Args:
        point_cloud_range (list[float]): Point cloud range.
    """
zhangwenwei's avatar
zhangwenwei committed
768
769

    def __init__(self, point_cloud_range):
770
        self.pcd_range = np.array(point_cloud_range, dtype=np.float32)
zhangwenwei's avatar
zhangwenwei committed
771
772

    def __call__(self, input_dict):
773
774
775
776
777
778
        """Call function to filter points by the range.

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

        Returns:
779
            dict: Results after filtering, 'points', 'pts_instance_mask'
780
                and 'pts_semantic_mask' keys are updated in the result dict.
781
        """
zhangwenwei's avatar
zhangwenwei committed
782
        points = input_dict['points']
783
784
        points_mask = points.in_range_3d(self.pcd_range)
        clean_points = points[points_mask]
zhangwenwei's avatar
zhangwenwei committed
785
        input_dict['points'] = clean_points
786
787
788
789
790
791
792
793
794
795
796
        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
797
798
799
        return input_dict

    def __repr__(self):
800
        """str: Return a string that describes the module."""
zhangwenwei's avatar
zhangwenwei committed
801
        repr_str = self.__class__.__name__
802
        repr_str += f'(point_cloud_range={self.pcd_range.tolist()})'
zhangwenwei's avatar
zhangwenwei committed
803
        return repr_str
zhangwenwei's avatar
zhangwenwei committed
804
805
806
807


@PIPELINES.register_module()
class ObjectNameFilter(object):
zhangwenwei's avatar
zhangwenwei committed
808
    """Filter GT objects by their names.
zhangwenwei's avatar
zhangwenwei committed
809
810

    Args:
liyinhao's avatar
liyinhao committed
811
        classes (list[str]): List of class names to be kept for training.
zhangwenwei's avatar
zhangwenwei committed
812
813
814
815
816
817
818
    """

    def __init__(self, classes):
        self.classes = classes
        self.labels = list(range(len(self.classes)))

    def __call__(self, input_dict):
819
820
821
822
823
824
        """Call function to filter objects by their names.

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

        Returns:
825
            dict: Results after filtering, 'gt_bboxes_3d', 'gt_labels_3d'
826
827
                keys are updated in the result dict.
        """
zhangwenwei's avatar
zhangwenwei committed
828
829
830
831
832
833
834
835
836
        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):
837
        """str: Return a string that describes the module."""
zhangwenwei's avatar
zhangwenwei committed
838
839
840
        repr_str = self.__class__.__name__
        repr_str += f'(classes={self.classes})'
        return repr_str
wuyuefeng's avatar
wuyuefeng committed
841
842
843


@PIPELINES.register_module()
844
845
class PointSample(object):
    """Point sample.
wuyuefeng's avatar
wuyuefeng committed
846
847
848
849
850

    Sampling data to a certain number.

    Args:
        num_points (int): Number of points to be sampled.
851
        sample_range (float, optional): The range where to sample points.
852
853
854
855
            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
856
857
    """

858
    def __init__(self, num_points, sample_range=None, replace=False):
wuyuefeng's avatar
wuyuefeng committed
859
        self.num_points = num_points
860
861
862
863
864
865
866
867
868
        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
869
870
871
872
873
        """Points random sampling.

        Sample points to a certain number.

        Args:
874
            points (np.ndarray | :obj:`BasePoints`): 3D Points.
wuyuefeng's avatar
wuyuefeng committed
875
            num_samples (int): Number of samples to be sampled.
876
            sample_range (float, optional): Indicating the range where the
877
                points will be sampled. Defaults to None.
878
879
880
881
            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
882
        Returns:
883
            tuple[np.ndarray] | np.ndarray:
884
                - points (np.ndarray | :obj:`BasePoints`): 3D Points.
885
                - choices (np.ndarray, optional): The generated random samples.
wuyuefeng's avatar
wuyuefeng committed
886
        """
887
        if not replace:
wuyuefeng's avatar
wuyuefeng committed
888
            replace = (points.shape[0] < num_samples)
889
890
891
892
893
894
        point_range = range(len(points))
        if sample_range is not None and not replace:
            # Only sampling the near points when len(points) >= num_samples
            depth = np.linalg.norm(points.tensor, axis=1)
            far_inds = np.where(depth > sample_range)[0]
            near_inds = np.where(depth <= sample_range)[0]
895
896
897
898
            # 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)
899
900
901
902
903
904
905
            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
906
907
908
909
910
911
        if return_choices:
            return points[choices], choices
        else:
            return points[choices]

    def __call__(self, results):
912
913
914
915
916
        """Call function to sample points to in indoor scenes.

        Args:
            input_dict (dict): Result dict from loading pipeline.
        Returns:
917
            dict: Results after sampling, 'points', 'pts_instance_mask'
918
919
                and 'pts_semantic_mask' keys are updated in the result dict.
        """
wuyuefeng's avatar
wuyuefeng committed
920
        points = results['points']
921
922
923
        # Points in Camera coord can provide the depth information.
        # TODO: Need to suport distance-based sampling for other coord system.
        if self.sample_range is not None:
924
            from mmdet3d.core.points import CameraPoints
925
926
927
928
929
930
931
932
            assert isinstance(points, CameraPoints), \
                'Sampling based on distance is only appliable for CAMERA coord'
        points, choices = self._points_random_sampling(
            points,
            self.num_points,
            self.sample_range,
            self.replace,
            return_choices=True)
933
        results['points'] = points
934

wuyuefeng's avatar
wuyuefeng committed
935
936
937
        pts_instance_mask = results.get('pts_instance_mask', None)
        pts_semantic_mask = results.get('pts_semantic_mask', None)

938
        if pts_instance_mask is not None:
wuyuefeng's avatar
wuyuefeng committed
939
940
            pts_instance_mask = pts_instance_mask[choices]
            results['pts_instance_mask'] = pts_instance_mask
941
942
943

        if pts_semantic_mask is not None:
            pts_semantic_mask = pts_semantic_mask[choices]
wuyuefeng's avatar
wuyuefeng committed
944
945
946
947
948
            results['pts_semantic_mask'] = pts_semantic_mask

        return results

    def __repr__(self):
949
        """str: Return a string that describes the module."""
wuyuefeng's avatar
wuyuefeng committed
950
        repr_str = self.__class__.__name__
951
        repr_str += f'(num_points={self.num_points},'
952
953
        repr_str += f' sample_range={self.sample_range},'
        repr_str += f' replace={self.replace})'
954

955
956
957
        return repr_str


958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
@PIPELINES.register_module()
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)


975
976
977
978
979
980
981
982
983
984
985
986
@PIPELINES.register_module()
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.
987
988
989
            This parameter is unused in `IndoorPatchPointSample` and thus has
            been deprecated. We plan to remove it in the future.
            Defaults to None.
990
991
        ignore_index (int, optional): Label index that won't be used for the
            segmentation task. This is set in PointSegClassMapping as neg_cls.
992
            If not None, will be used as a patch selection criterion.
993
994
995
996
997
            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.
998
        enlarge_size (float, optional): Enlarge the sampled patch to
999
            [-block_size / 2 - enlarge_size, block_size / 2 + enlarge_size] as
1000
            an augmentation. If None, set it as 0. Defaults to 0.2.
1001
        min_unique_num (int, optional): Minimum number of unique points
1002
1003
            the sampled patch should contain. If None, use PointNet++'s method
            to judge uniqueness. Defaults to None.
1004
1005
        eps (float, optional): A value added to patch boundary to guarantee
            points coverage. Defaults to 1e-2.
1006
1007
1008
1009
1010
1011

    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.
1012
1013
1014
1015
1016
    """

    def __init__(self,
                 num_points,
                 block_size=1.5,
1017
                 sample_rate=None,
1018
1019
                 ignore_index=None,
                 use_normalized_coord=False,
1020
1021
                 num_try=10,
                 enlarge_size=0.2,
1022
1023
                 min_unique_num=None,
                 eps=1e-2):
1024
1025
1026
1027
1028
        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
1029
        self.enlarge_size = enlarge_size if enlarge_size is not None else 0.0
1030
        self.min_unique_num = min_unique_num
1031
        self.eps = eps
1032
1033
1034
1035
1036

        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.')
1037
1038
1039
1040
1041

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

1042
        Generate input by subtracting patch center and adding additional
1043
1044
1045
1046
1047
1048
1049
1050
1051
            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.
1052
            point_type (type): class of input points inherited from BasePoints.
1053
1054

        Returns:
1055
            :obj:`BasePoints`: The generated input data.
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
        """
        # 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

1079
    def _patch_points_sampling(self, points, sem_mask):
1080
1081
1082
1083
1084
1085
        """Patch points sampling.

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

        Args:
1086
            points (:obj:`BasePoints`): 3D Points.
1087
1088
1089
            sem_mask (np.ndarray): semantic segmentation mask for input points.

        Returns:
1090
            tuple[:obj:`BasePoints`, np.ndarray] | :obj:`BasePoints`:
1091

1092
                - points (:obj:`BasePoints`): 3D Points.
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
                - 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)

1103
        for _ in range(self.num_try):
1104
1105
1106
            # random sample a point as patch center
            cur_center = coords[np.random.choice(coords.shape[0])]

1107
1108
            # boundary of a patch, which would be enlarged by
            # `self.enlarge_size` as an augmentation
1109
1110
1111
1112
1113
1114
1115
            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(
1116
1117
                (coords >= (cur_min - self.enlarge_size)) *
                (coords <= (cur_max + self.enlarge_size)),
1118
1119
1120
1121
1122
1123
1124
                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]
1125
            point_idxs = np.where(cur_choice)[0]
1126
            mask = np.sum(
1127
1128
                (cur_coords >= (cur_min - self.eps)) * (cur_coords <=
                                                        (cur_max + self.eps)),
1129
                axis=1) == 3
1130

1131
1132
            # two criteria for patch sampling, adopted from PointNet++
            # 1. selected patch should contain enough unique points
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
            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:
1145
                # if `min_unique_num` is provided, directly compare with it
1146
                flag1 = mask.sum() >= self.min_unique_num
1147

1148
            # 2. selected patch should contain enough annotated points
1149
1150
1151
1152
1153
1154
1155
1156
1157
            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

1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
        # 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]
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185

        # 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:
1186
            dict: Results after sampling, 'points', 'pts_instance_mask'
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
                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},'
1213
1214
        repr_str += f' num_try={self.num_try},'
        repr_str += f' enlarge_size={self.enlarge_size},'
1215
1216
        repr_str += f' min_unique_num={self.min_unique_num},'
        repr_str += f' eps={self.eps})'
wuyuefeng's avatar
wuyuefeng committed
1217
        return repr_str
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245


@PIPELINES.register_module()
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:
1246
            dict: Results after filtering, 'points', 'pts_instance_mask'
1247
                and 'pts_semantic_mask' keys are updated in the result dict.
1248
1249
1250
1251
        """
        points = input_dict['points']
        gt_bboxes_3d = input_dict['gt_bboxes_3d']

xiliu8006's avatar
xiliu8006 committed
1252
1253
1254
1255
        # 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()

1256
1257
        enlarged_gt_bboxes_3d = gt_bboxes_3d_np.copy()
        enlarged_gt_bboxes_3d[:, 3:6] += self.bbox_enlarge_range
xiliu8006's avatar
xiliu8006 committed
1258
        points_numpy = points.tensor.clone().numpy()
1259
1260
        foreground_masks = box_np_ops.points_in_rbbox(
            points_numpy, gt_bboxes_3d_np, origin=(0.5, 0.5, 0.5))
1261
        enlarge_foreground_masks = box_np_ops.points_in_rbbox(
1262
            points_numpy, enlarged_gt_bboxes_3d, origin=(0.5, 0.5, 0.5))
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
        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__
1281
        repr_str += f'(bbox_enlarge_range={self.bbox_enlarge_range.tolist()})'
1282
        return repr_str
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343


@PIPELINES.register_module()
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.
        time_dim (int): Index that indicate the time dimention
            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.
            point_dim (int): The dimention of each points

        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:
1344
            dict: Results after sampling, 'points', 'pts_instance_mask'
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
                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
1355
1356
        points_numpy = points.tensor.numpy()
        extra_channel = [points_numpy]
1357
1358
1359
1360
1361
1362
1363
1364
1365
        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])

1366
        points_numpy = np.concatenate(extra_channel, axis=-1)
1367
1368
1369
1370
1371

        # 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.
1372
1373
1374
        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]
1375
1376
1377
1378
1379
1380
1381
1382
1383
        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,
1384
                                               points_numpy.shape[1])
1385
1386
1387
        if self.prev_voxel_generator is not None:
            prev_sweeps_points = self._sample_points(prev_sweeps_points,
                                                     self.prev_voxel_generator,
1388
                                                     points_numpy.shape[1])
1389

1390
1391
            points_numpy = np.concatenate(
                [cur_sweep_points, prev_sweeps_points], 0)
1392
        else:
1393
            points_numpy = cur_sweep_points
1394
1395

        if self.cur_voxel_generator._max_num_points == 1:
1396
1397
            points_numpy = points_numpy.squeeze(1)
        results['points'] = points.new_point(points_numpy[..., :original_dim])
1398
1399
1400

        # Restore the correspoinding seg and mask fields
        for key, dim_index in map_fields2dim:
1401
            results[key] = points_numpy[..., dim_index]
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424

        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