-
Notifications
You must be signed in to change notification settings - Fork 0
/
pascal_parts.py
238 lines (190 loc) · 9.77 KB
/
pascal_parts.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
"""
Code adapted from: https://github.com/akanazawa/cmr/blob/master/data/cub.py
MIT License
Copyright (c) 2018 akanazawa
"""
import os
import cv2
import scipy.io
from tqdm import tqdm
cv2.setNumThreads(0)
import numpy as np
import torchvision
from PIL import Image
from torch.utils.data import Dataset
from torchvision.transforms import functional as TF, InterpolationMode
from utils import image as image_utils
def pil_loader(path, type):
with open(path, 'rb') as f:
img = Image.open(f)
return img.convert(type)
def pad_if_smaller(img, size, fill=None):
min_size = min(img.shape[:2])
if min_size < size:
ow, oh = img.shape[:2]
padh = size - oh if oh < size else 0
padw = size - ow if ow < size else 0
pad = ((padw // 2, padw - padw // 2), (padh // 2, padh - padh // 2), (0,0)) if len(img.shape) == 3 else ((padw // 2, padw - padw // 2), (padh // 2, padh - padh // 2))
if fill is None:
img = np.pad(img, pad, 'edge')
else:
img = np.pad(img, pad, 'constant', constant_values=fill)
return img
def bbox2(img):
rows = np.any(img, axis=1)
cols = np.any(img, axis=0)
rmin, rmax = np.where(rows)[0][[0, -1]]
cmin, cmax = np.where(cols)[0][[0, -1]]
return [cmin, rmin, cmax, rmax]
def pct_area(img, bbox):
x0, y0, x1, y1 = bbox
image_size = img.shape
return (x1 - x0) * (y1 - y0) / (image_size[0] * image_size[1] + 1e-7)
padding_frac = 0.05
jitter_frac = 0.05
dict_part = {'tvmonitor': ['background', 'screen'],
'cat': ['background', 'head', 'lbleg', 'lbpa', 'lear', 'leye', 'lfleg', 'lfpa', 'neck', 'nose', 'rbleg', 'rbpa', 'rear', 'reye', 'rfleg', 'rfpa', 'tail', 'torso'],
'person': ['background', 'hair', 'head', 'lear', 'lebrow', 'leye', 'lfoot', 'lhand', 'llarm', 'llleg', 'luarm', 'luleg', 'mouth', 'neck', 'nose', 'rear', 'rebrow', 'reye', 'rfoot', 'rhand', 'rlarm', 'rlleg', 'ruarm', 'ruleg', 'torso'],
'motorbike': ['background', 'bwheel', 'fwheel', 'handlebar', 'headlight', 'saddle'],
'car': ['background', 'backside', 'bliplate', 'door', 'fliplate', 'frontside', 'headlight', 'leftmirror', 'leftside', 'rightmirror', 'rightside', 'roofside', 'wheel', 'window'],
'aeroplane': ['background', 'body', 'engine', 'lwing', 'rwing', 'stern', 'tail', 'wheel'],
'dog': ['background', 'head', 'lbleg', 'lbpa', 'lear', 'leye', 'lfleg', 'lfpa', 'muzzle', 'neck', 'nose', 'rbleg', 'rbpa', 'rear', 'reye', 'rfleg', 'rfpa', 'tail', 'torso'],
'bus': ['background', 'backside', 'bliplate', 'door', 'fliplate', 'frontside', 'headlight', 'leftmirror', 'leftside', 'rightmirror', 'rightside', 'roofside', 'wheel', 'window'],
'train': ['background', 'cbackside', 'cfrontside', 'cleftside', 'coach', 'crightside', 'croofside', 'hbackside', 'head', 'headlight', 'hfrontside', 'hleftside', 'hrightside', 'hroofside'],
'bird': ['background', 'beak', 'head', 'leye', 'lfoot', 'lleg', 'lwing', 'neck', 'reye', 'rfoot', 'rleg', 'rwing', 'tail', 'torso'],
'horse': ['background', 'head', 'lbho', 'lblleg', 'lbuleg', 'lear', 'leye', 'lfho', 'lflleg', 'lfuleg', 'muzzle', 'neck', 'rbho', 'rblleg', 'rbuleg', 'rear', 'reye', 'rfho', 'rflleg', 'rfuleg', 'tail', 'torso'],
'pottedplant': ['background', 'plant', 'pot'],
'cow': ['background', 'head', 'lblleg', 'lbuleg', 'lear', 'leye', 'lflleg', 'lfuleg', 'lhorn', 'muzzle', 'neck', 'rblleg', 'rbuleg', 'rear', 'reye', 'rflleg', 'rfuleg', 'rhorn', 'tail', 'torso'],
'bicycle': ['background', 'bwheel', 'chainwheel', 'fwheel', 'handlebar', 'headlight', 'saddle'],
'bottle': ['background', 'body', 'cap'],
'sheep': ['background', 'head', 'lblleg', 'lbuleg', 'lear', 'leye', 'lflleg', 'lfuleg', 'lhorn', 'muzzle', 'neck', 'rblleg', 'rbuleg', 'rear', 'reye', 'rflleg', 'rfuleg', 'rhorn', 'tail', 'torso']}
# https://github.com/micco00x/py-pascalpart
# Load annotations from .mat files creating a Python dictionary:
def load_annotations(path):
# Get annotations from the file and relative objects:
annotations = scipy.io.loadmat(path)["anno"]
objects = annotations[0, 0]["objects"]
# List containing information of each object (to add to dictionary):
objects_list = []
# Go through the objects and extract info:
for obj_idx in range(objects.shape[1]):
obj = objects[0, obj_idx]
# Get classname and mask of the current object:
classname = obj["class"][0]
mask = obj["mask"]
# List containing information of each body part (to add to dictionary):
parts_list = []
parts = obj["parts"]
# Go through the part of the specific object and extract info:
for part_idx in range(parts.shape[1]):
part = parts[0, part_idx]
# Get part name and mask of the current body part:
part_name = part["part_name"][0]
part_mask = part["mask"]
# Add info to parts_list:
parts_list.append({"part_name": part_name, "mask": part_mask})
# Add info to objects_list:
objects_list.append({"class": classname, "mask": mask, "parts": parts_list})
return {"objects": objects_list}
class PPDataset(Dataset):
def __init__(self, opts):
super().__init__()
self.opts = opts
self.img_size = 224
self.split = "train"
self.dataset_root = opts.data_dir
self.dataset = 'pascal-parts'
self.jitter_frac = jitter_frac
self.padding_frac = padding_frac
split = "train"
self.masks = []
self.images = []
self.bbox = []
annotation_folder = f'{self.dataset_root}/Annotations_Part/'
images_folder = f'{self.dataset_root}/VOC2010/JPEGImages/'
cls = opts.pascal_class
mat_filenames = os.listdir(annotation_folder)
voc_list = {str(s) for s in np.loadtxt(f'{self.dataset_root}/VOC2010/ImageSets/Main/{cls}_{"train" if split == "train" else "val"}.txt', dtype=str)[:, 0]}
for idx, annotation_filename in enumerate(tqdm(mat_filenames)):
if annotation_filename.split('.')[0] in voc_list:
annotations = load_annotations(os.path.join(annotation_folder, annotation_filename))
for obj in annotations["objects"]:
if obj["class"] == cls:
bbox = bbox2(obj['mask'])
mask = np.zeros_like(obj['mask'])
# if pct_area(obj['mask'], bbox) > (0.20 if split == 'test' else 0.10):
for body_part in obj["parts"][::-1]:
instance_mask = body_part["mask"].astype(np.uint8)
part_name = body_part["part_name"].split('_')[0]
mask = mask * (1 - instance_mask) + instance_mask * dict_part[obj["class"]].index(part_name)
self.bbox.append(bbox)
self.images.append(images_folder + annotation_filename[:annotation_filename.rfind(".")] + ".jpg")
self.masks.append(mask.astype(np.uint8))
print(f"Total {split}: {len(voc_list)} {cls} {split}: {len(self.images)}")
@staticmethod
def only_file_names(lst):
return [e['file_name'] for e in lst]
def forward_img(self, index):
path = self.images[index]
img = pil_loader(path, 'RGB')
mask = self.masks[index]
mask = Image.fromarray(mask)
img = np.array(img)
mask = np.array(mask)
# Some are grayscale:
if len(img.shape) == 2:
img = np.repeat(np.expand_dims(img, 2), 3, axis=2)
mask = np.expand_dims(mask, 2)
h, w, _ = mask.shape
bbox = self.bbox[index]
if self.split == 'train':
bbox = image_utils.peturb_bbox(bbox, pf=self.padding_frac, jf=self.jitter_frac)
else:
bbox = image_utils.peturb_bbox(bbox, pf=self.padding_frac, jf=0)
bbox = image_utils.square_bbox(bbox)
img, mask = self.crop_image(img, mask, bbox)
# scale image, and mask. And scale kps.
img, mask = self.scale_image(img, mask)
# Mirror image on random.
if self.split == 'train':
img, mask = self.mirror_image(img, mask)
img = Image.fromarray(img.astype(np.uint8))
mask = np.asarray(mask, np.uint8)
return img, mask, path
def crop_image(self, img, mask, bbox):
# crop image and mask and translate kps
img = image_utils.crop(img, bbox, bgval=1)
mask = image_utils.crop(mask, bbox, bgval=0)
return img, mask
def scale_image(self, img, mask):
# Scale image so largest bbox size is img_size
bwidth = np.shape(img)[0]
bheight = np.shape(img)[1]
scale = self.img_size / float(max(bwidth, bheight))
img_scale, _ = image_utils.resize_img(img, scale)
mask_scale, _ = image_utils.resize_img(mask, scale)
mask_scale = np.expand_dims(mask_scale, 2)
img_scale = pad_if_smaller(img_scale, self.img_size)
mask_scale = pad_if_smaller(mask_scale, self.img_size)
return img_scale, mask_scale
def mirror_image(self, img, mask):
if np.random.rand(1) > 0.5:
# Need copy bc torch collate doesnt like neg strides
img_flip = img[:, ::-1, :].copy()
mask_flip = mask[:, ::-1].copy()
return img_flip, mask_flip
else:
return img, mask
def __len__(self):
return len(self.images)
def __getitem__(self, index):
img, seg, img_path = self.forward_img(index)
mask = (seg != 0).astype(np.uint8)
elem = {
'img': img,
'mask': mask,
'seg': seg,
'inds': index,
'img_path': img_path,
}
return elem