File size: 10,149 Bytes
34d1f8b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
# Copyright (c) OpenMMLab. All rights reserved.
import os
from concurrent import futures as futures
from os import path as osp

import mmengine
import numpy as np


class S3DISData(object):
    """S3DIS data.

    Generate s3dis infos for s3dis_converter.

    Args:
        root_path (str): Root path of the raw data.
        split (str, optional): Set split type of the data. Default: 'Area_1'.
    """

    def __init__(self, root_path, split='Area_1'):
        self.root_dir = root_path
        self.split = split
        self.data_dir = osp.join(root_path,
                                 'Stanford3dDataset_v1.2_Aligned_Version')

        # Following `GSDN <https://arxiv.org/abs/2006.12356>`_, use 5 furniture
        # classes for detection: table, chair, sofa, bookcase, board.
        self.cat_ids = np.array([7, 8, 9, 10, 11])
        self.cat_ids2class = {
            cat_id: i
            for i, cat_id in enumerate(list(self.cat_ids))
        }

        assert split in [
            'Area_1', 'Area_2', 'Area_3', 'Area_4', 'Area_5', 'Area_6'
        ]
        self.sample_id_list = os.listdir(osp.join(self.data_dir,
                                                  split))  # conferenceRoom_1
        for sample_id in self.sample_id_list:
            if os.path.isfile(osp.join(self.data_dir, split, sample_id)):
                self.sample_id_list.remove(sample_id)

    def __len__(self):
        return len(self.sample_id_list)

    def get_infos(self, num_workers=4, has_label=True, sample_id_list=None):
        """Get data infos.

        This method gets information from the raw data.

        Args:
            num_workers (int, optional): Number of threads to be used.
                Default: 4.
            has_label (bool, optional): Whether the data has label.
                Default: True.
            sample_id_list (list[int], optional): Index list of the sample.
                Default: None.

        Returns:
            infos (list[dict]): Information of the raw data.
        """

        def process_single_scene(sample_idx):
            print(f'{self.split} sample_idx: {sample_idx}')
            info = dict()
            pc_info = {
                'num_features': 6,
                'lidar_idx': f'{self.split}_{sample_idx}'
            }
            info['point_cloud'] = pc_info
            pts_filename = osp.join(self.root_dir, 's3dis_data',
                                    f'{self.split}_{sample_idx}_point.npy')
            pts_instance_mask_path = osp.join(
                self.root_dir, 's3dis_data',
                f'{self.split}_{sample_idx}_ins_label.npy')
            pts_semantic_mask_path = osp.join(
                self.root_dir, 's3dis_data',
                f'{self.split}_{sample_idx}_sem_label.npy')

            points = np.load(pts_filename).astype(np.float32)
            pts_instance_mask = np.load(pts_instance_mask_path).astype(
                np.int64)
            pts_semantic_mask = np.load(pts_semantic_mask_path).astype(
                np.int64)

            mmengine.mkdir_or_exist(osp.join(self.root_dir, 'points'))
            mmengine.mkdir_or_exist(osp.join(self.root_dir, 'instance_mask'))
            mmengine.mkdir_or_exist(osp.join(self.root_dir, 'semantic_mask'))

            points.tofile(
                osp.join(self.root_dir, 'points',
                         f'{self.split}_{sample_idx}.bin'))
            pts_instance_mask.tofile(
                osp.join(self.root_dir, 'instance_mask',
                         f'{self.split}_{sample_idx}.bin'))
            pts_semantic_mask.tofile(
                osp.join(self.root_dir, 'semantic_mask',
                         f'{self.split}_{sample_idx}.bin'))

            info['pts_path'] = osp.join('points',
                                        f'{self.split}_{sample_idx}.bin')
            info['pts_instance_mask_path'] = osp.join(
                'instance_mask', f'{self.split}_{sample_idx}.bin')
            info['pts_semantic_mask_path'] = osp.join(
                'semantic_mask', f'{self.split}_{sample_idx}.bin')
            info['annos'] = self.get_bboxes(points, pts_instance_mask,
                                            pts_semantic_mask)

            return info

        sample_id_list = sample_id_list if sample_id_list is not None \
            else self.sample_id_list
        with futures.ThreadPoolExecutor(num_workers) as executor:
            infos = executor.map(process_single_scene, sample_id_list)
        return list(infos)

    def get_bboxes(self, points, pts_instance_mask, pts_semantic_mask):
        """Convert instance masks to axis-aligned bounding boxes.

        Args:
            points (np.array): Scene points of shape (n, 6).
            pts_instance_mask (np.ndarray): Instance labels of shape (n,).
            pts_semantic_mask (np.ndarray): Semantic labels of shape (n,).

        Returns:
            dict: A dict containing detection infos with following keys:

                - gt_boxes_upright_depth (np.ndarray): Bounding boxes
                    of shape (n, 6)
                - class (np.ndarray): Box labels of shape (n,)
                - gt_num (int): Number of boxes.
        """
        bboxes, labels = [], []
        for i in range(1, pts_instance_mask.max() + 1):
            ids = pts_instance_mask == i
            # check if all instance points have same semantic label
            assert pts_semantic_mask[ids].min() == pts_semantic_mask[ids].max()
            label = pts_semantic_mask[ids][0]
            # keep only furniture objects
            if label in self.cat_ids2class:
                labels.append(self.cat_ids2class[pts_semantic_mask[ids][0]])
                pts = points[:, :3][ids]
                min_pts = pts.min(axis=0)
                max_pts = pts.max(axis=0)
                locations = (min_pts + max_pts) / 2
                dimensions = max_pts - min_pts
                bboxes.append(np.concatenate((locations, dimensions)))
        annotation = dict()
        # follow ScanNet and SUN RGB-D keys
        annotation['gt_boxes_upright_depth'] = np.array(bboxes)
        annotation['class'] = np.array(labels)
        annotation['gt_num'] = len(labels)
        return annotation


class S3DISSegData(object):
    """S3DIS dataset used to generate infos for semantic segmentation task.

    Args:
        data_root (str): Root path of the raw data.
        ann_file (str): The generated scannet infos.
        split (str, optional): Set split type of the data. Default: 'train'.
        num_points (int, optional): Number of points in each data input.
            Default: 8192.
        label_weight_func (function, optional): Function to compute the
            label weight. Default: None.
    """

    def __init__(self,
                 data_root,
                 ann_file,
                 split='Area_1',
                 num_points=4096,
                 label_weight_func=None):
        self.data_root = data_root
        self.data_infos = mmengine.load(ann_file)
        self.split = split
        self.num_points = num_points

        self.all_ids = np.arange(13)  # all possible ids
        self.cat_ids = np.array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
                                 12])  # used for seg task
        self.ignore_index = len(self.cat_ids)

        self.cat_id2class = np.ones(
            (self.all_ids.shape[0], ), dtype=np.int64) * self.ignore_index
        for i, cat_id in enumerate(self.cat_ids):
            self.cat_id2class[cat_id] = i

        # label weighting function is taken from
        # https://github.com/charlesq34/pointnet2/blob/master/scannet/scannet_dataset.py#L24
        self.label_weight_func = (lambda x: 1.0 / np.log(1.2 + x)) if \
            label_weight_func is None else label_weight_func

    def get_seg_infos(self):
        scene_idxs, label_weight = self.get_scene_idxs_and_label_weight()
        save_folder = osp.join(self.data_root, 'seg_info')
        mmengine.mkdir_or_exist(save_folder)
        np.save(
            osp.join(save_folder, f'{self.split}_resampled_scene_idxs.npy'),
            scene_idxs)
        np.save(
            osp.join(save_folder, f'{self.split}_label_weight.npy'),
            label_weight)
        print(f'{self.split} resampled scene index and label weight saved')

    def _convert_to_label(self, mask):
        """Convert class_id in loaded segmentation mask to label."""
        if isinstance(mask, str):
            if mask.endswith('npy'):
                mask = np.load(mask)
            else:
                mask = np.fromfile(mask, dtype=np.int64)
        label = self.cat_id2class[mask]
        return label

    def get_scene_idxs_and_label_weight(self):
        """Compute scene_idxs for data sampling and label weight for loss
        calculation.

        We sample more times for scenes with more points. Label_weight is
        inversely proportional to number of class points.
        """
        num_classes = len(self.cat_ids)
        num_point_all = []
        label_weight = np.zeros((num_classes + 1, ))  # ignore_index
        for data_info in self.data_infos:
            label = self._convert_to_label(
                osp.join(self.data_root, data_info['pts_semantic_mask_path']))
            num_point_all.append(label.shape[0])
            class_count, _ = np.histogram(label, range(num_classes + 2))
            label_weight += class_count

        # repeat scene_idx for num_scene_point // num_sample_point times
        sample_prob = np.array(num_point_all) / float(np.sum(num_point_all))
        num_iter = int(np.sum(num_point_all) / float(self.num_points))
        scene_idxs = []
        for idx in range(len(self.data_infos)):
            scene_idxs.extend([idx] * int(round(sample_prob[idx] * num_iter)))
        scene_idxs = np.array(scene_idxs).astype(np.int32)

        # calculate label weight, adopted from PointNet++
        label_weight = label_weight[:-1].astype(np.float32)
        label_weight = label_weight / label_weight.sum()
        label_weight = self.label_weight_func(label_weight).astype(np.float32)

        return scene_idxs, label_weight