-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathdataset.py
executable file
·244 lines (193 loc) · 10 KB
/
dataset.py
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
from video_records import DataEgo_VideoRecord, CrossDataEgo_VideoRecord, MMAct_VideoRecord, mmdata_VideoRecord
import torch.utils.data as data
from PIL import Image
import os
import os.path
import pandas as pd
import numpy as np
from numpy.random import randint
import pickle
class MMTSADataSet(data.Dataset):
def __init__(self, dataset, list_file,
new_length, modality, image_tmpl,
visual_path=None, sensor_path=None,
num_segments=3, transform=None,
mode='train', cross_dataset = False):
self.dataset = dataset
self.visual_path = visual_path
self.list_file = list_file
self.num_segments = num_segments
self.new_length = new_length
self.modality = modality
self.image_tmpl = image_tmpl
self.transform = transform
self.mode = mode
self.cross_dataset = cross_dataset
self._parse_list()
def _GramianAngularField(self, series, fps = 15.0):
image_size = series.shape[1]
from pyts.image import GramianAngularField
gasf = GramianAngularField(image_size=image_size, method='summation')
sensor_gasf = gasf.fit_transform(series)
return sensor_gasf
def _normalization(self, data, scale = 255.0):
_range = np.max(data) - np.min(data)
return (data - np.min(data)) / _range * 255.0
def _extract_sensor_feature(self, record, idx):
# 确定中间秒
centre_sec = (record.start_frame + idx) / record.fps['Sensor']
# 左右各1s
left_sec = centre_sec - 1.0
right_sec = centre_sec + 1.0
# sensor数据 (行数 x 6个channel)
sensor_data = np.load(record.sensor_path, allow_pickle=True).astype('float')[:,:6]
duration = sensor_data.shape[0] / float(record.fps['Sensor'])
left_sample = int(round(left_sec * record.fps['Sensor']))
right_sample = int(round(right_sec * record.fps['Sensor']))
if left_sec < 0:
samples = sensor_data[:int(round(record.fps['Sensor'] * 2.0))]
elif right_sec > duration:
samples = sensor_data[-int(round(record.fps['Sensor'] * 2.0)):]
else:
samples = sensor_data[left_sample:right_sample]
return self._GramianAngularField(samples.transpose(), record.fps['Sensor'])
def _extract_accphone_feature(self, record, idx):
centre_sec = (record.start_frame + idx) / record.fps['AccPhone']
left_sec = centre_sec - 1.0
right_sec = centre_sec + 1.0
sensor_data = np.load(record.AccPhone_path, allow_pickle=True).astype('float')[:,:3]
duration = sensor_data.shape[0] / float(record.fps['AccPhone'])
left_sample = int(round(left_sec * record.fps['AccPhone']))
right_sample = int(round(right_sec * record.fps['AccPhone']))
if left_sec < 0:
samples = sensor_data[:int(round(record.fps['AccPhone'] * 2.0))]
elif right_sec > duration or right_sample > sensor_data.shape[0]:
samples = sensor_data[-int(round(record.fps['AccPhone'] * 2.0)):]
else:
samples = sensor_data[left_sample:right_sample]
return self._GramianAngularField(samples.transpose(), record.fps['AccPhone'])
def _extract_accwatch_feature(self, record, idx):
centre_sec = (record.start_frame + idx) / record.fps['AccWatch']
left_sec = centre_sec - 1.0
right_sec = centre_sec + 1.0
sensor_data = np.load(record.AccWatch_path, allow_pickle=True).astype('float')[:,:3]
duration = sensor_data.shape[0] / float(record.fps['AccWatch'])
left_sample = int(round(left_sec * record.fps['AccWatch']))
right_sample = int(round(right_sec * record.fps['AccWatch']))
if left_sec < 0:
samples = sensor_data[:int(round(record.fps['AccWatch'] * 2.0))]
elif right_sec > duration or right_sample > sensor_data.shape[0]:
samples = sensor_data[-int(round(record.fps['AccWatch'] * 2.0)):]
else:
samples = sensor_data[left_sample:right_sample]
return self._GramianAngularField(samples.transpose(), record.fps['AccWatch'])
def _extract_gyro_feature(self, record, idx):
centre_sec = (record.start_frame + idx) / record.fps['Gyro']
left_sec = centre_sec - 1.0
right_sec = centre_sec + 1.0
sensor_data = np.load(record.Gyro_path, allow_pickle=True).astype('float')[:,:3]
duration = sensor_data.shape[0] / float(record.fps['Gyro'])
left_sample = int(round(left_sec * record.fps['Gyro']))
right_sample = int(round(right_sec * record.fps['Gyro']))
if left_sec < 0:
samples = sensor_data[:int(round(record.fps['Gyro'] * 2.0))]
elif right_sec > duration or right_sample > sensor_data.shape[0]:
samples = sensor_data[-int(round(record.fps['Gyro'] * 2.0)):]
else:
samples = sensor_data[left_sample:right_sample]
return self._GramianAngularField(samples.transpose(), record.fps['Gyro'])
def _extract_orie_feature(self, record, idx):
centre_sec = (record.start_frame + idx) / record.fps['Orie']
left_sec = centre_sec - 1.0
right_sec = centre_sec + 1.0
sensor_data = np.load(record.Orie_path, allow_pickle=True).astype('float')[:,:3]
duration = sensor_data.shape[0] / float(record.fps['Orie'])
left_sample = int(round(left_sec * record.fps['Orie']))
right_sample = int(round(right_sec * record.fps['Orie']))
if left_sec < 0:
samples = sensor_data[:int(round(record.fps['Orie'] * 2.0))]
elif right_sec > duration or right_sample > sensor_data.shape[0]:
samples = sensor_data[-int(round(record.fps['Orie'] * 2.0)):]
else:
samples = sensor_data[left_sample:right_sample]
return self._GramianAngularField(samples.transpose(), record.fps['Orie'])
def _load_data(self, modality, record, idx):
if self.dataset == 'MMAct':
video_path = record.video_path
else:
video_path = os.path.join(self.visual_path, record.untrimmed_video_name)
if modality == 'RGB':
idx_untrimmed = record.start_frame + idx
if idx_untrimmed==0:
idx_untrimmed += 1
return [Image.open(os.path.join(video_path, self.image_tmpl[modality].format(idx_untrimmed))).convert('RGB')]
elif modality =="Sensor":
sens = self._extract_sensor_feature(record, idx)
return [Image.fromarray(self._normalization(single_channel)).convert('L') for single_channel in sens]
elif modality =="AccPhone":
sens = self._extract_accphone_feature(record, idx)
return [Image.fromarray(self._normalization(single_channel)).convert('L') for single_channel in sens]
elif modality =="AccWatch":
sens = self._extract_accwatch_feature(record, idx)
return [Image.fromarray(self._normalization(single_channel)).convert('L') for single_channel in sens]
elif modality =="Gyro":
sens = self._extract_gyro_feature(record, idx)
return [Image.fromarray(self._normalization(single_channel)).convert('L') for single_channel in sens]
elif modality =="Orie":
sens = self._extract_orie_feature(record, idx)
return [Image.fromarray(self._normalization(single_channel)).convert('L') for single_channel in sens]
def _parse_list(self):
if self.dataset == 'dataEgo':
if self.cross_dataset == False:
self.video_list = [DataEgo_VideoRecord(tup) for tup in self.list_file.iterrows()]
else:
self.video_list = [CrossDataEgo_VideoRecord(tup) for tup in self.list_file.iterrows()]
elif self.dataset == 'MMAct':
self.video_list = [MMAct_VideoRecord(tup) for tup in self.list_file.iterrows()]
elif self.dataset == 'mmdata':
self.video_list = [mmdata_VideoRecord(tup) for tup in self.list_file.iterrows()]
def _sample_indices(self, record, modality):
"""
:param record: VideoRecord
:return: list
"""
average_duration = (record.num_frames[modality] - self.new_length[modality] + 1) // self.num_segments
if average_duration > 0:
offsets = np.multiply(list(range(self.num_segments)), average_duration) + randint(average_duration, size=self.num_segments)
else:
offsets = np.zeros((self.num_segments,))
return offsets
def _get_val_indices(self, record, modality):
if record.num_frames[modality] > self.num_segments + self.new_length[modality] - 1:
tick = (record.num_frames[modality] - self.new_length[modality] + 1) / float(self.num_segments)
offsets = np.array([int(tick / 2.0 + tick * x) for x in range(self.num_segments)])
else:
offsets = np.zeros((self.num_segments,))
return offsets
def __getitem__(self, index):
input = {}
record = self.video_list[index]
for m in self.modality:
if self.mode == 'train':
segment_indices = self._sample_indices(record, m)
else:
segment_indices = self._get_val_indices(record, m)
if m != 'RGB' and self.mode == 'train':
np.random.shuffle(segment_indices)
img, label = self.get(m, record, segment_indices)
input[m] = img
# print(index, input['RGB'].shape, input['Sensor'].shape)
return input, label
def get(self, modality, record, indices):
images = list()
for seg_ind in indices:
p = int(seg_ind)
for i in range(self.new_length[modality]):
seg_imgs = self._load_data(modality, record, p)
images.extend(seg_imgs)
if p < record.num_frames[modality]:
p += 1
process_data = self.transform[modality](images)
return process_data, int(record.label)
def __len__(self):
return len(self.video_list)