det3d_dataset.py 16.7 KB
Newer Older
jshilong's avatar
jshilong committed
1
2
# Copyright (c) OpenMMLab. All rights reserved.
import copy
3
import os
jshilong's avatar
jshilong committed
4
from os import path as osp
5
from typing import Callable, List, Optional, Set, Union
jshilong's avatar
jshilong committed
6

7
import mmengine
jshilong's avatar
jshilong committed
8
import numpy as np
9
import torch
jshilong's avatar
jshilong committed
10
from mmengine.dataset import BaseDataset
11
12
from mmengine.logging import print_log
from terminaltables import AsciiTable
jshilong's avatar
jshilong committed
13
14

from mmdet3d.datasets import DATASETS
zhangshilong's avatar
zhangshilong committed
15
from mmdet3d.structures import get_box_type
jshilong's avatar
jshilong committed
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31


@DATASETS.register_module()
class Det3DDataset(BaseDataset):
    """Base Class of 3D dataset.

    This is the base dataset of SUNRGB-D, ScanNet, nuScenes, and KITTI
    dataset.
    # TODO: doc link here for the standard data format

    Args:
        data_root (str, optional): The root directory for ``data_prefix`` and
            ``ann_file``. Defaults to None.
        ann_file (str): Annotation file path. Defaults to ''.
        metainfo (dict, optional): Meta information for dataset, such as class
            information. Defaults to None.
32
        data_prefix (dict): Prefix for training data. Defaults to
33
            dict(pts='velodyne', img='').
34
        pipeline (List[dict]): Pipeline used for data processing.
35
36
37
            Defaults to [].
        modality (dict): Modality to specify the sensor data used as input,
            it usually has following keys:
jshilong's avatar
jshilong committed
38
39
40

                - use_camera: bool
                - use_lidar: bool
41
            Defaults to dict(use_lidar=True, use_camera=False).
jshilong's avatar
jshilong committed
42
43
        default_cam_key (str, optional): The default camera name adopted.
            Defaults to None.
44
        box_type_3d (str): Type of 3D box of this dataset.
jshilong's avatar
jshilong committed
45
46
            Based on the `box_type_3d`, the dataset will encapsulate the box
            to its original format then converted them to `box_type_3d`.
47
            Defaults to 'LiDAR' in this dataset. Available options includes:
jshilong's avatar
jshilong committed
48
49
50
51
52
53
54

            - 'LiDAR': Box in LiDAR coordinates, usually for
              outdoor point cloud 3d detection.
            - 'Depth': Box in depth coordinates, usually for
              indoor point cloud 3d detection.
            - 'Camera': Box in camera coordinates, usually
              for vision-based 3d detection.
55
56
57
58
59
        filter_empty_gt (bool): Whether to filter the data with empty GT.
            If it's set to be True, the example with empty annotations after
            data pipeline will be dropped and a random example will be chosen
            in `__getitem__`. Defaults to True.
        test_mode (bool): Whether the dataset is in test mode.
jshilong's avatar
jshilong committed
60
            Defaults to False.
61
62
63
64
        load_eval_anns (bool): Whether to load annotations in test_mode,
            the annotation will be save in `eval_ann_infos`, which can be
            used in Evaluator. Defaults to True.
        file_client_args (dict): Configuration of file client.
65
            Defaults to dict(backend='disk').
66
67
68
        show_ins_var (bool): For debug purpose. Whether to show variation
            of the number of instances before and after through pipeline.
            Defaults to False.
jshilong's avatar
jshilong committed
69
70
71
72
73
74
75
76
77
    """

    def __init__(self,
                 data_root: Optional[str] = None,
                 ann_file: str = '',
                 metainfo: Optional[dict] = None,
                 data_prefix: dict = dict(pts='velodyne', img=''),
                 pipeline: List[Union[dict, Callable]] = [],
                 modality: dict = dict(use_lidar=True, use_camera=False),
jshilong's avatar
jshilong committed
78
                 default_cam_key: str = None,
jshilong's avatar
jshilong committed
79
80
81
                 box_type_3d: dict = 'LiDAR',
                 filter_empty_gt: bool = True,
                 test_mode: bool = False,
82
                 load_eval_anns: bool = True,
jshilong's avatar
jshilong committed
83
                 file_client_args: dict = dict(backend='disk'),
84
                 show_ins_var: bool = False,
85
                 **kwargs) -> None:
jshilong's avatar
jshilong committed
86
        # init file client
87
        self.file_client = mmengine.FileClient(**file_client_args)
jshilong's avatar
jshilong committed
88
        self.filter_empty_gt = filter_empty_gt
jshilong's avatar
jshilong committed
89
        self.load_eval_anns = load_eval_anns
jshilong's avatar
jshilong committed
90
91
92
93
94
95
96
97
98
        _default_modality_keys = ('use_lidar', 'use_camera')
        if modality is None:
            modality = dict()

        # Defaults to False if not specify
        for key in _default_modality_keys:
            if key not in modality:
                modality[key] = False
        self.modality = modality
jshilong's avatar
jshilong committed
99
        self.default_cam_key = default_cam_key
jshilong's avatar
jshilong committed
100
101
        assert self.modality['use_lidar'] or self.modality['use_camera'], (
            'Please specify the `modality` (`use_lidar` '
jshilong's avatar
jshilong committed
102
            f', `use_camera`) for {self.__class__.__name__}')
jshilong's avatar
jshilong committed
103
104

        self.box_type_3d, self.box_mode_3d = get_box_type(box_type_3d)
VVsssssk's avatar
VVsssssk committed
105

106
107
        if metainfo is not None and 'classes' in metainfo:
            # we allow to train on subset of self.METAINFO['classes']
jshilong's avatar
jshilong committed
108
109
110
            # map unselected labels to -1
            self.label_mapping = {
                i: -1
111
                for i in range(len(self.METAINFO['classes']))
jshilong's avatar
jshilong committed
112
113
            }
            self.label_mapping[-1] = -1
114
115
            for label_idx, name in enumerate(metainfo['classes']):
                ori_label = self.METAINFO['classes'].index(name)
jshilong's avatar
jshilong committed
116
                self.label_mapping[ori_label] = label_idx
117

118
            self.num_ins_per_cat = {name: 0 for name in metainfo['classes']}
jshilong's avatar
jshilong committed
119
120
121
        else:
            self.label_mapping = {
                i: i
122
                for i in range(len(self.METAINFO['classes']))
jshilong's avatar
jshilong committed
123
124
125
            }
            self.label_mapping[-1] = -1

126
127
            self.num_ins_per_cat = {
                name: 0
128
                for name in self.METAINFO['classes']
129
130
            }

jshilong's avatar
jshilong committed
131
132
133
134
135
136
137
138
139
        super().__init__(
            ann_file=ann_file,
            metainfo=metainfo,
            data_root=data_root,
            data_prefix=data_prefix,
            pipeline=pipeline,
            test_mode=test_mode,
            **kwargs)

VVsssssk's avatar
VVsssssk committed
140
141
142
143
        # can be accessed by other component in runner
        self.metainfo['box_type_3d'] = box_type_3d
        self.metainfo['label_mapping'] = self.label_mapping

144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
        # used for showing variation of the number of instances before and
        # after through the pipeline
        self.show_ins_var = show_ins_var

        # show statistics of this dataset
        print_log('-' * 30, 'current')
        print_log(f'The length of the dataset: {len(self)}', 'current')
        content_show = [['category', 'number']]
        for cat_name, num in self.num_ins_per_cat.items():
            content_show.append([cat_name, num])
        table = AsciiTable(content_show)
        print_log(
            f'The number of instances per category in the dataset:\n{table.table}',  # noqa: E501
            'current')

159
    def _remove_dontcare(self, ann_info: dict) -> dict:
jshilong's avatar
jshilong committed
160
161
        """Remove annotations that do not need to be cared.

162
        -1 indicates dontcare in MMDet3d.
jshilong's avatar
jshilong committed
163
164
165
166
167
168
169
170
171
172
173

        Args:
            ann_info (dict): Dict of annotation infos. The
                instance with label `-1` will be removed.

        Returns:
            dict: Annotations after filtering.
        """
        img_filtered_annotations = {}
        filter_mask = ann_info['gt_labels_3d'] > -1
        for key in ann_info.keys():
zhangshilong's avatar
zhangshilong committed
174
175
176
177
            if key != 'instances':
                img_filtered_annotations[key] = (ann_info[key][filter_mask])
            else:
                img_filtered_annotations[key] = ann_info[key]
jshilong's avatar
jshilong committed
178
179
180
181
182
183
184
185
186
187
188
189
        return img_filtered_annotations

    def get_ann_info(self, index: int) -> dict:
        """Get annotation info according to the given index.

        Use index to get the corresponding annotations, thus the
        evalhook could use this api.

        Args:
            index (int): Index of the annotation data to get.

        Returns:
190
            dict: Annotation information.
jshilong's avatar
jshilong committed
191
192
193
194
195
196
197
198
199
200
        """
        data_info = self.get_data_info(index)
        # test model
        if 'ann_info' not in data_info:
            ann_info = self.parse_ann_info(data_info)
        else:
            ann_info = data_info['ann_info']

        return ann_info

201
    def parse_ann_info(self, info: dict) -> Union[dict, None]:
202
        """Process the `instances` in data info to `ann_info`.
jshilong's avatar
jshilong committed
203
204
205
206
207
208
209
210
211
212

        In `Custom3DDataset`, we simply concatenate all the field
        in `instances` to `np.ndarray`, you can do the specific
        process in subclass. You have to convert `gt_bboxes_3d`
        to different coordinates according to the task.

        Args:
            info (dict): Info dict.

        Returns:
213
            dict or None: Processed `ann_info`.
jshilong's avatar
jshilong committed
214
215
        """
        # add s or gt prefix for most keys after concat
zhangshilong's avatar
zhangshilong committed
216
217
        # we only process 3d annotations here, the corresponding
        # 2d annotation process is in the `LoadAnnotations3D`
zhangshilong's avatar
zhangshilong committed
218
        # in `transforms`
jshilong's avatar
jshilong committed
219
220
        name_mapping = {
            'bbox_label_3d': 'gt_labels_3d',
221
222
            'bbox_label': 'gt_bboxes_labels',
            'bbox': 'gt_bboxes',
jshilong's avatar
jshilong committed
223
224
225
            'bbox_3d': 'gt_bboxes_3d',
            'depth': 'depths',
            'center_2d': 'centers_2d',
ChaimZhu's avatar
ChaimZhu committed
226
227
            'attr_label': 'attr_labels',
            'velocity': 'velocities',
jshilong's avatar
jshilong committed
228
229
        }
        instances = info['instances']
230
231
232
233
234
235
236
        # empty gt
        if len(instances) == 0:
            return None
        else:
            keys = list(instances[0].keys())
            ann_info = dict()
            for ann_name in keys:
zhangshilong's avatar
zhangshilong committed
237
238
                temp_anns = [item[ann_name] for item in instances]
                # map the original dataset label to training label
239
                if 'label' in ann_name and ann_name != 'attr_label':
zhangshilong's avatar
zhangshilong committed
240
241
242
                    temp_anns = [
                        self.label_mapping[item] for item in temp_anns
                    ]
243
                if ann_name in name_mapping:
ChaimZhu's avatar
ChaimZhu committed
244
245
246
                    mapped_ann_name = name_mapping[ann_name]
                else:
                    mapped_ann_name = ann_name
247
248
249

                if 'label' in ann_name:
                    temp_anns = np.array(temp_anns).astype(np.int64)
ChaimZhu's avatar
ChaimZhu committed
250
                elif ann_name in name_mapping:
251
                    temp_anns = np.array(temp_anns).astype(np.float32)
ChaimZhu's avatar
ChaimZhu committed
252
253
                else:
                    temp_anns = np.array(temp_anns)
254

ChaimZhu's avatar
ChaimZhu committed
255
                ann_info[mapped_ann_name] = temp_anns
zhangshilong's avatar
zhangshilong committed
256
            ann_info['instances'] = info['instances']
257
258

            for label in ann_info['gt_labels_3d']:
259
260
261
                if label != -1:
                    cat_name = self.metainfo['classes'][label]
                    self.num_ins_per_cat[cat_name] += 1
262

jshilong's avatar
jshilong committed
263
264
265
266
267
268
        return ann_info

    def parse_data_info(self, info: dict) -> dict:
        """Process the raw data info.

        Convert all relative path of needed modality data file to
269
270
        the absolute path. And process the `instances` field to
        `ann_info` in training stage.
jshilong's avatar
jshilong committed
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285

        Args:
            info (dict): Raw info dict.

        Returns:
            dict: Has `ann_info` in training stage. And
            all path has been converted to absolute path.
        """

        if self.modality['use_lidar']:
            info['lidar_points']['lidar_path'] = \
                osp.join(
                    self.data_prefix.get('pts', ''),
                    info['lidar_points']['lidar_path'])

ChaimZhu's avatar
ChaimZhu committed
286
            info['num_pts_feats'] = info['lidar_points']['num_pts_feats']
jshilong's avatar
jshilong committed
287
            info['lidar_path'] = info['lidar_points']['lidar_path']
VVsssssk's avatar
VVsssssk committed
288
289
290
            if 'lidar_sweeps' in info:
                for sweep in info['lidar_sweeps']:
                    file_suffix = sweep['lidar_points']['lidar_path'].split(
291
                        os.sep)[-1]
VVsssssk's avatar
VVsssssk committed
292
293
294
295
296
297
                    if 'samples' in sweep['lidar_points']['lidar_path']:
                        sweep['lidar_points']['lidar_path'] = osp.join(
                            self.data_prefix['pts'], file_suffix)
                    else:
                        sweep['lidar_points']['lidar_path'] = osp.join(
                            self.data_prefix['sweeps'], file_suffix)
jshilong's avatar
jshilong committed
298

jshilong's avatar
jshilong committed
299
300
301
        if self.modality['use_camera']:
            for cam_id, img_info in info['images'].items():
                if 'img_path' in img_info:
VVsssssk's avatar
VVsssssk committed
302
303
304
305
306
307
                    if cam_id in self.data_prefix:
                        cam_prefix = self.data_prefix[cam_id]
                    else:
                        cam_prefix = self.data_prefix.get('img', '')
                    img_info['img_path'] = osp.join(cam_prefix,
                                                    img_info['img_path'])
jshilong's avatar
jshilong committed
308
309
310
311
312
313
314
315
316
317
318
319
320
321
            if self.default_cam_key is not None:
                info['img_path'] = info['images'][
                    self.default_cam_key]['img_path']
                if 'lidar2cam' in info['images'][self.default_cam_key]:
                    info['lidar2cam'] = np.array(
                        info['images'][self.default_cam_key]['lidar2cam'])
                if 'cam2img' in info['images'][self.default_cam_key]:
                    info['cam2img'] = np.array(
                        info['images'][self.default_cam_key]['cam2img'])
                if 'lidar2img' in info['images'][self.default_cam_key]:
                    info['lidar2img'] = np.array(
                        info['images'][self.default_cam_key]['lidar2img'])
                else:
                    info['lidar2img'] = info['cam2img'] @ info['lidar2cam']
jshilong's avatar
jshilong committed
322
323

        if not self.test_mode:
Tai-Wang's avatar
Tai-Wang committed
324
            # used in training
jshilong's avatar
jshilong committed
325
            info['ann_info'] = self.parse_ann_info(info)
jshilong's avatar
jshilong committed
326
327
        if self.test_mode and self.load_eval_anns:
            info['eval_ann_info'] = self.parse_ann_info(info)
jshilong's avatar
jshilong committed
328
329
330

        return info

331
332
    def _show_ins_var(self, old_labels: np.ndarray,
                      new_labels: torch.Tensor) -> None:
333
334
335
336
337
338
339
340
341
        """Show variation of the number of instances before and after through
        the pipeline.

        Args:
            old_labels (np.ndarray): The labels before through the pipeline.
            new_labels (torch.Tensor): The labels after through the pipeline.
        """
        ori_num_per_cat = dict()
        for label in old_labels:
342
343
344
345
            if label != -1:
                cat_name = self.metainfo['classes'][label]
                ori_num_per_cat[cat_name] = ori_num_per_cat.get(cat_name,
                                                                0) + 1
346
347
        new_num_per_cat = dict()
        for label in new_labels:
348
349
350
351
            if label != -1:
                cat_name = self.metainfo['classes'][label]
                new_num_per_cat[cat_name] = new_num_per_cat.get(cat_name,
                                                                0) + 1
352
353
354
355
356
357
358
359
360
        content_show = [['category', 'new number', 'ori number']]
        for cat_name, num in ori_num_per_cat.items():
            new_num = new_num_per_cat.get(cat_name, 0)
            content_show.append([cat_name, new_num, num])
        table = AsciiTable(content_show)
        print_log(
            'The number of instances per category after and before '
            f'through pipeline:\n{table.table}', 'current')

361
    def prepare_data(self, index: int) -> Union[dict, None]:
jshilong's avatar
jshilong committed
362
363
364
365
366
367
368
369
        """Data preparation for both training and testing stage.

        Called by `__getitem__`  of dataset.

        Args:
            index (int): Index for accessing the target data.

        Returns:
370
            dict or None: Data dict of the corresponding index.
jshilong's avatar
jshilong committed
371
        """
372
        ori_input_dict = self.get_data_info(index)
jshilong's avatar
jshilong committed
373
374

        # deepcopy here to avoid inplace modification in pipeline.
375
        input_dict = copy.deepcopy(ori_input_dict)
jshilong's avatar
jshilong committed
376
377
378
379
380
381
382
383
384
385
386
387

        # box_type_3d (str): 3D box type.
        input_dict['box_type_3d'] = self.box_type_3d
        # box_mode_3d (str): 3D box mode.
        input_dict['box_mode_3d'] = self.box_mode_3d

        # pre-pipline return None to random another in `__getitem__`
        if not self.test_mode and self.filter_empty_gt:
            if len(input_dict['ann_info']['gt_labels_3d']) == 0:
                return None

        example = self.pipeline(input_dict)
388

jshilong's avatar
jshilong committed
389
390
391
        if not self.test_mode and self.filter_empty_gt:
            # after pipeline drop the example with empty annotations
            # return None to random another in `__getitem__`
392
            if example is None or len(
393
                    example['data_samples'].gt_instances_3d.labels_3d) == 0:
jshilong's avatar
jshilong committed
394
                return None
395
396

        if self.show_ins_var:
397
398
399
400
401
402
403
404
405
406
            if 'ann_info' in ori_input_dict:
                self._show_ins_var(
                    ori_input_dict['ann_info']['gt_labels_3d'],
                    example['data_samples'].gt_instances_3d.labels_3d)
            else:
                print_log(
                    "'ann_info' is not in the input dict. It's probably that "
                    'the data is not in training mode',
                    'current',
                    level=30)
407

jshilong's avatar
jshilong committed
408
        return example
409

410
    def get_cat_ids(self, idx: int) -> Set[int]:
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
        """Get category ids by index. Dataset wrapped by ClassBalancedDataset
        must implement this method.

        The ``CBGSDataset`` or ``ClassBalancedDataset``requires a subclass
        which implements this method.

        Args:
            idx (int): The index of data.

        Returns:
            set[int]: All categories in the sample of specified index.
        """
        info = self.get_data_info(idx)
        gt_labels = info['ann_info']['gt_labels_3d'].tolist()
        return set(gt_labels)