dbsampler.py 11.7 KB
Newer Older
dingchang's avatar
dingchang committed
1
# Copyright (c) OpenMMLab. All rights reserved.
zhangwenwei's avatar
zhangwenwei committed
2
import copy
3
4
import os

5
import mmcv
zhangwenwei's avatar
zhangwenwei committed
6
import numpy as np
zhangwenwei's avatar
zhangwenwei committed
7
8
9

from mmdet3d.core.bbox import box_np_ops
from mmdet3d.datasets.pipelines import data_augment_utils
10
from mmdet.datasets import PIPELINES
11
from ..builder import OBJECTSAMPLERS
zhangwenwei's avatar
zhangwenwei committed
12
13
14


class BatchSampler:
wangtai's avatar
wangtai committed
15
16
17
18
    """Class for sampling specific category of ground truths.

    Args:
        sample_list (list[dict]): List of samples.
19
20
21
22
        name (str, optional): The category of samples. Default: None.
        epoch (int, optional): Sampling epoch. Default: None.
        shuffle (bool, optional): Whether to shuffle indices. Default: False.
        drop_reminder (bool, optional): Drop reminder. Default: False.
wangtai's avatar
wangtai committed
23
    """
zhangwenwei's avatar
zhangwenwei committed
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43

    def __init__(self,
                 sampled_list,
                 name=None,
                 epoch=None,
                 shuffle=True,
                 drop_reminder=False):
        self._sampled_list = sampled_list
        self._indices = np.arange(len(sampled_list))
        if shuffle:
            np.random.shuffle(self._indices)
        self._idx = 0
        self._example_num = len(sampled_list)
        self._name = name
        self._shuffle = shuffle
        self._epoch = epoch
        self._epoch_counter = 0
        self._drop_reminder = drop_reminder

    def _sample(self, num):
wangtai's avatar
wangtai committed
44
45
46
47
48
49
50
51
        """Sample specific number of ground truths and return indices.

        Args:
            num (int): Sampled number.

        Returns:
            list[int]: Indices of sampled ground truths.
        """
zhangwenwei's avatar
zhangwenwei committed
52
53
54
55
56
57
58
59
60
        if self._idx + num >= self._example_num:
            ret = self._indices[self._idx:].copy()
            self._reset()
        else:
            ret = self._indices[self._idx:self._idx + num]
            self._idx += num
        return ret

    def _reset(self):
wangtai's avatar
wangtai committed
61
        """Reset the index of batchsampler to zero."""
zhangwenwei's avatar
zhangwenwei committed
62
63
64
65
66
67
68
        assert self._name is not None
        # print("reset", self._name)
        if self._shuffle:
            np.random.shuffle(self._indices)
        self._idx = 0

    def sample(self, num):
wangtai's avatar
wangtai committed
69
70
71
72
73
74
75
76
        """Sample specific number of ground truths.

        Args:
            num (int): Sampled number.

        Returns:
            list[dict]: Sampled ground truths.
        """
zhangwenwei's avatar
zhangwenwei committed
77
78
79
80
        indices = self._sample(num)
        return [self._sampled_list[i] for i in indices]


81
@OBJECTSAMPLERS.register_module()
zhangwenwei's avatar
zhangwenwei committed
82
class DataBaseSampler(object):
83
84
85
86
87
88
89
90
    """Class for sampling data from the ground truth database.

    Args:
        info_path (str): Path of groundtruth database info.
        data_root (str): Path of groundtruth database.
        rate (float): Rate of actual sampled over maximum sampled number.
        prepare (dict): Name of preparation functions and the input value.
        sample_groups (dict): Sampled classes and numbers.
91
92
93
        classes (list[str], optional): List of classes. Default: None.
        points_loader(dict, optional): Config of points loader. Default:
            dict(type='LoadPointsFromFile', load_dim=4, use_dim=[0,1,2,3])
94
    """
zhangwenwei's avatar
zhangwenwei committed
95

zhangwenwei's avatar
zhangwenwei committed
96
97
98
99
100
101
    def __init__(self,
                 info_path,
                 data_root,
                 rate,
                 prepare,
                 sample_groups,
102
103
104
                 classes=None,
                 points_loader=dict(
                     type='LoadPointsFromFile',
105
                     coord_type='LIDAR',
106
107
                     load_dim=4,
                     use_dim=[0, 1, 2, 3])):
zhangwenwei's avatar
zhangwenwei committed
108
        super().__init__()
zhangwenwei's avatar
zhangwenwei committed
109
        self.data_root = data_root
zhangwenwei's avatar
zhangwenwei committed
110
111
112
        self.info_path = info_path
        self.rate = rate
        self.prepare = prepare
zhangwenwei's avatar
zhangwenwei committed
113
114
115
        self.classes = classes
        self.cat2label = {name: i for i, name in enumerate(classes)}
        self.label2cat = {i: name for i, name in enumerate(classes)}
116
        self.points_loader = mmcv.build_from_cfg(points_loader, PIPELINES)
zhangwenwei's avatar
zhangwenwei committed
117

118
        db_infos = mmcv.load(info_path)
zhangwenwei's avatar
zhangwenwei committed
119
120

        # filter database infos
zhangwenwei's avatar
zhangwenwei committed
121
        from mmdet3d.utils import get_root_logger
zhangwenwei's avatar
zhangwenwei committed
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
        logger = get_root_logger()
        for k, v in db_infos.items():
            logger.info(f'load {len(v)} {k} database infos')
        for prep_func, val in prepare.items():
            db_infos = getattr(self, prep_func)(db_infos, val)
        logger.info('After filter database:')
        for k, v in db_infos.items():
            logger.info(f'load {len(v)} {k} database infos')

        self.db_infos = db_infos

        # load sample groups
        # TODO: more elegant way to load sample groups
        self.sample_groups = []
        for name, num in sample_groups.items():
            self.sample_groups.append({name: int(num)})

        self.group_db_infos = self.db_infos  # just use db_infos
        self.sample_classes = []
        self.sample_max_nums = []
        for group_info in self.sample_groups:
            self.sample_classes += list(group_info.keys())
            self.sample_max_nums += list(group_info.values())

        self.sampler_dict = {}
        for k, v in self.group_db_infos.items():
            self.sampler_dict[k] = BatchSampler(v, k, shuffle=True)
        # TODO: No group_sampling currently

    @staticmethod
    def filter_by_difficulty(db_infos, removed_difficulty):
153
154
155
156
157
158
159
160
161
        """Filter ground truths by difficulties.

        Args:
            db_infos (dict): Info of groundtruth database.
            removed_difficulty (list): Difficulties that are not qualified.

        Returns:
            dict: Info of database after filtering.
        """
zhangwenwei's avatar
zhangwenwei committed
162
163
164
165
166
167
168
169
170
171
        new_db_infos = {}
        for key, dinfos in db_infos.items():
            new_db_infos[key] = [
                info for info in dinfos
                if info['difficulty'] not in removed_difficulty
            ]
        return new_db_infos

    @staticmethod
    def filter_by_min_points(db_infos, min_gt_points_dict):
172
173
174
175
176
177
178
179
180
181
        """Filter ground truths by number of points in the bbox.

        Args:
            db_infos (dict): Info of groundtruth database.
            min_gt_points_dict (dict): Different number of minimum points
                needed for different categories of ground truths.

        Returns:
            dict: Info of database after filtering.
        """
zhangwenwei's avatar
zhangwenwei committed
182
183
184
185
186
187
188
189
190
191
        for name, min_num in min_gt_points_dict.items():
            min_num = int(min_num)
            if min_num > 0:
                filtered_infos = []
                for info in db_infos[name]:
                    if info['num_points_in_gt'] >= min_num:
                        filtered_infos.append(info)
                db_infos[name] = filtered_infos
        return db_infos

192
    def sample_all(self, gt_bboxes, gt_labels, img=None, ground_plane=None):
193
194
195
196
        """Sampling all categories of bboxes.

        Args:
            gt_bboxes (np.ndarray): Ground truth bounding boxes.
liyinhao's avatar
liyinhao committed
197
            gt_labels (np.ndarray): Ground truth labels of boxes.
198
199
200
201

        Returns:
            dict: Dict of sampled 'pseudo ground truths'.

202
                - gt_labels_3d (np.ndarray): ground truths labels
zhangwenwei's avatar
zhangwenwei committed
203
                    of sampled objects.
204
                - gt_bboxes_3d (:obj:`BaseInstance3DBoxes`):
liyinhao's avatar
liyinhao committed
205
                    sampled ground truth 3D bounding boxes
206
207
208
                - points (np.ndarray): sampled points
                - group_ids (np.ndarray): ids of sampled ground truths
        """
zhangwenwei's avatar
zhangwenwei committed
209
210
211
212
        sampled_num_dict = {}
        sample_num_per_class = []
        for class_name, max_sample_num in zip(self.sample_classes,
                                              self.sample_max_nums):
zhangwenwei's avatar
zhangwenwei committed
213
214
215
            class_label = self.cat2label[class_name]
            # sampled_num = int(max_sample_num -
            #                   np.sum([n == class_name for n in gt_names]))
zhangwenwei's avatar
zhangwenwei committed
216
            sampled_num = int(max_sample_num -
zhangwenwei's avatar
zhangwenwei committed
217
                              np.sum([n == class_label for n in gt_labels]))
zhangwenwei's avatar
zhangwenwei committed
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
            sampled_num = np.round(self.rate * sampled_num).astype(np.int64)
            sampled_num_dict[class_name] = sampled_num
            sample_num_per_class.append(sampled_num)

        sampled = []
        sampled_gt_bboxes = []
        avoid_coll_boxes = gt_bboxes

        for class_name, sampled_num in zip(self.sample_classes,
                                           sample_num_per_class):
            if sampled_num > 0:
                sampled_cls = self.sample_class_v2(class_name, sampled_num,
                                                   avoid_coll_boxes)

                sampled += sampled_cls
                if len(sampled_cls) > 0:
                    if len(sampled_cls) == 1:
                        sampled_gt_box = sampled_cls[0]['box3d_lidar'][
                            np.newaxis, ...]
                    else:
                        sampled_gt_box = np.stack(
                            [s['box3d_lidar'] for s in sampled_cls], axis=0)

                    sampled_gt_bboxes += [sampled_gt_box]
                    avoid_coll_boxes = np.concatenate(
                        [avoid_coll_boxes, sampled_gt_box], axis=0)

        ret = None
        if len(sampled) > 0:
            sampled_gt_bboxes = np.concatenate(sampled_gt_bboxes, axis=0)
            # center = sampled_gt_bboxes[:, 0:3]

zhangwenwei's avatar
zhangwenwei committed
250
            # num_sampled = len(sampled)
zhangwenwei's avatar
zhangwenwei committed
251
252
253
254
            s_points_list = []
            count = 0
            for info in sampled:
                file_path = os.path.join(
zhangwenwei's avatar
zhangwenwei committed
255
256
                    self.data_root,
                    info['path']) if self.data_root else info['path']
257
258
                results = dict(pts_filename=file_path)
                s_points = self.points_loader(results)['points']
259
                s_points.translate(info['box3d_lidar'][:3])
zhangwenwei's avatar
zhangwenwei committed
260
261
262
263

                count += 1

                s_points_list.append(s_points)
264
265
266

            gt_labels = np.array([self.cat2label[s['name']] for s in sampled],
                                 dtype=np.long)
267
268
269
270
271
272
273
274
275

            if ground_plane is not None:
                xyz = sampled_gt_bboxes[:, :3]
                dz = (ground_plane[:3][None, :] *
                      xyz).sum(-1) + ground_plane[3]
                sampled_gt_bboxes[:, 2] -= dz
                for i, s_points in enumerate(s_points_list):
                    s_points.tensor[:, 2].sub_(dz[i])

zhangwenwei's avatar
zhangwenwei committed
276
            ret = {
zhangwenwei's avatar
zhangwenwei committed
277
278
                'gt_labels_3d':
                gt_labels,
zhangwenwei's avatar
zhangwenwei committed
279
280
281
                'gt_bboxes_3d':
                sampled_gt_bboxes,
                'points':
282
                s_points_list[0].cat(s_points_list),
zhangwenwei's avatar
zhangwenwei committed
283
284
285
286
287
288
289
290
                'group_ids':
                np.arange(gt_bboxes.shape[0],
                          gt_bboxes.shape[0] + len(sampled))
            }

        return ret

    def sample_class_v2(self, name, num, gt_bboxes):
291
292
293
294
295
296
297
298
299
300
        """Sampling specific categories of bounding boxes.

        Args:
            name (str): Class of objects to be sampled.
            num (int): Number of sampled bboxes.
            gt_bboxes (np.ndarray): Ground truth boxes.

        Returns:
            list[dict]: Valid samples after collision test.
        """
zhangwenwei's avatar
zhangwenwei committed
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
        sampled = self.sampler_dict[name].sample(num)
        sampled = copy.deepcopy(sampled)
        num_gt = gt_bboxes.shape[0]
        num_sampled = len(sampled)
        gt_bboxes_bv = box_np_ops.center_to_corner_box2d(
            gt_bboxes[:, 0:2], gt_bboxes[:, 3:5], gt_bboxes[:, 6])

        sp_boxes = np.stack([i['box3d_lidar'] for i in sampled], axis=0)
        boxes = np.concatenate([gt_bboxes, sp_boxes], axis=0).copy()

        sp_boxes_new = boxes[gt_bboxes.shape[0]:]
        sp_boxes_bv = box_np_ops.center_to_corner_box2d(
            sp_boxes_new[:, 0:2], sp_boxes_new[:, 3:5], sp_boxes_new[:, 6])

        total_bv = np.concatenate([gt_bboxes_bv, sp_boxes_bv], axis=0)
        coll_mat = data_augment_utils.box_collision_test(total_bv, total_bv)
        diag = np.arange(total_bv.shape[0])
        coll_mat[diag, diag] = False

        valid_samples = []
        for i in range(num_gt, num_gt + num_sampled):
            if coll_mat[i].any():
                coll_mat[i] = False
                coll_mat[:, i] = False
            else:
                valid_samples.append(sampled[i - num_gt])
        return valid_samples