-
Notifications
You must be signed in to change notification settings - Fork 1
/
dataloader_vcoco.py
168 lines (148 loc) · 7.71 KB
/
dataloader_vcoco.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
from __future__ import print_function, division
import json
import os
import torch
import pickle
import random
import pandas as pd
import numpy as np
from PIL import Image
import data_preprocess as dp
from skimage import io, transform
from utils import processing_sg, LIS
from transforms import build_transforms
from torchvision import transforms, utils
from structures.bounding_box import BoxList
from torch.utils.data import Dataset, DataLoader
from structures.image_list import to_image_list
# Image list ids that cannot be detected any objects.
bad_detections_train = [6357, 17487, 130851, 153892, 176148, 326601, 364400, 365013, 369213, 429514, 483039, 560726]
bad_detections_val = [130099, 196981, 352877, 387895, 426849, 526087, 567439, 568117]
bad_detections_test = [16875, 30828, 56701, 66706, 75768, 165157, 228418, 230501, 255483, 267725, 293855, 344045,
413805, 419143, 451038, 479280, 561411]
def vcoco_collate(batch):
transposed_batch = list(zip(*batch))
images = to_image_list(transposed_batch[0], 32)
targets = transposed_batch[1]
return images, targets
class Rescale(object):
def __init__(self, output_size):
assert isinstance(output_size, (int, tuple))
self.output_size = output_size
def __call__(self, image):
h, w = image.shape[:2]
if isinstance(self.output_size, int):
if h > w:
new_h, new_w = self.output_size * h / w, self.output_size
else:
new_h, new_w = self.output_size, self.output_size * w / h
else:
new_h, new_w = self.output_size
new_h, new_w = int(new_h), int(new_w)
transformed_img = transform.resize(image, (new_h, new_w))
return transformed_img
class VcocoDataset(Dataset):
def __init__(self, json_file_image, root_dir, cfg):
with open(json_file_image) as json_file_:
self.vcoco_frame_file = json.load(json_file_)
self.flag = json_file_image.split('/')[-1].split('_')[0]
self.cfg = cfg
if self.flag == 'train':
self.vcoco_frame = [x for x in self.vcoco_frame_file.keys() if x not in str(bad_detections_train)]
self.detect_results = pickle.load(open(cfg.train_detected_results, 'rb'))
gt_path = cfg.data_dir + 'Annotations_vcoco/train_annotations.json'
with open(gt_path) as fp:
self.annotations = json.load(fp)
elif self.flag == 'val':
self.vcoco_frame = [x for x in self.vcoco_frame_file.keys() if x not in str(bad_detections_val)]
self.detect_results = pickle.load(open(cfg.train_detected_results, 'rb'))
gt_path = cfg.data_dir + 'Annotations_vcoco/val_annotations.json'
with open(gt_path) as fp:
self.annotations = json.load(fp)
elif self.flag == 'test':
self.vcoco_frame = [x for x in self.vcoco_frame_file.keys() if x not in str(bad_detections_test)]
self.detect_results = pickle.load(open(cfg.test_detected_results, 'rb'))
gt_path = cfg.data_dir + 'Annotations_vcoco/test_annotations.json'
with open(gt_path) as fp:
self.annotations = json.load(fp)
self.root_dir = root_dir
self.sg_pred = pickle.load(open(cfg.sg_data + self.flag + '_0.4_0.2.pk', 'rb'))
self.transform = build_transforms(cfg, is_train=(self.flag == 'train'))
self.max_nagetive = 512
def __len__(self):
return len(self.vcoco_frame)
def convert2target(self, image, res):
img_info = res['shape']
w, h = img_info[0], img_info[1]
box = np.concatenate((res['per_box'], res['obj_box']))
scales = [w, h, w, h]
box2 = box / scales
box = torch.from_numpy(box).clone()
target = BoxList(box, (w, h), 'xyxy') # xyxy
if self.transform is not None:
image, target = self.transform(image, target)
target.bbox[res['per_box'].shape[0], :] = 0
per_obj_labs = np.concatenate((np.array([1 for _ in range(res['per_box'].shape[0])]), res['all_obj_labels']))
target.add_field("labels", torch.from_numpy(per_obj_labs))
target.add_field('boxes', torch.from_numpy(box2))
per_scores = torch.tensor(res['scores_persons']).float()
per_scores = LIS(per_scores, 8.3, 12, 10)
obj_scores = torch.tensor(res['scores_objects']).float()
obj_scores = LIS(obj_scores, 8.3, 12, 10)
labels_scores = torch.cat((per_scores, obj_scores))
target.add_field('obj_scores', labels_scores)
pair_score = []
pair_info = []
hoi_labs = res['labels_all']
target.add_field('hoi_labels', torch.from_numpy(hoi_labs))
all_labels = target.get_field('labels')[res['labels_all'].shape[0]:]
obj_label_for_mask = []
one_hot_labs = []
num_bg = 0
for i in range(res['labels_all'].shape[0]):
for j in range(res['labels_all'].shape[1]):
if self.flag == 'test':
pair_info.append([i, res['labels_all'].shape[0] + j, 1])
one_hot_labs.append(target.get_field('hoi_labels')[i, j, :])
pair_score.append(per_scores[i] * obj_scores[j])
obj_label_for_mask.append(all_labels[j])
else:
if int(res['labels_all'][i, j, :].sum()) > 0:
pair_info.append([i, res['labels_all'].shape[0] + j, 1])
one_hot_labs.append(target.get_field('hoi_labels')[i, j, :])
pair_score.append(per_scores[i] * obj_scores[j])
obj_label_for_mask.append(all_labels[j])
elif num_bg < self.max_nagetive:
pair_info.append([i, res['labels_all'].shape[0] + j, 0])
one_hot_labs.append(target.get_field('hoi_labels')[i, j, :])
pair_score.append(per_scores[i] * obj_scores[j])
num_bg += 1
obj_label_for_mask.append(all_labels[j])
elif random.random() < 0.5:
replace_id = int(random.random() * len(pair_info))
pair_info[replace_id] = [i, res['labels_all'].shape[0] + j, 0]
one_hot_labs[replace_id] = target.get_field('hoi_labels')[i, j, :]
pair_score[replace_id] = per_scores[i] * obj_scores[j]
obj_label_for_mask[replace_id] = all_labels[j]
target.add_field("pairs_info", torch.tensor(pair_info))
target.add_field("mask", torch.tensor(obj_label_for_mask))
target.add_field("per_mul_obj_scores", torch.tensor(pair_score))
target.add_field('HOI_labs', torch.stack(one_hot_labs))
return image, target
def __getitem__(self, idx):
if self.flag == 'test':
img_pre_suffix = 'COCO_val2014_' + str(self.vcoco_frame[idx]).zfill(12) + '.jpg'
else:
img_pre_suffix = 'COCO_train2014_' + str(self.vcoco_frame[idx]).zfill(12) + '.jpg'
img_name = os.path.join(self.root_dir, img_pre_suffix)
image = Image.open(img_name).convert('RGB')
all_info = dp.get_anotation_info_by_imageId(int(self.vcoco_frame[idx]), self.flag, self.detect_results,
self.annotations)
image, target = self.convert2target(image, all_info)
union_box = dp.get_attention_maps(target)
target.add_field('union_box', torch.tensor(union_box).float())
target.add_field('sg', self.sg_pred[img_pre_suffix])
sg_graph = processing_sg(self.sg_pred[img_pre_suffix], target)
target.add_field('sg_data', sg_graph)
target.add_field('image_id', self.vcoco_frame[idx])
return image, target