File size: 9,342 Bytes
3a1265d |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 |
"""
HICODet dataset under PyTorch framework
Fred Zhang <frederic.zhang@anu.edu.au>
The Australian National University
Australian Centre for Robotic Vision
"""
import os
import json
import numpy as np
from typing import Optional, List, Callable, Tuple
from pocket.data import ImageDataset, DataSubset
class HICODetSubset(DataSubset):
def __init__(self, *args) -> None:
super().__init__(*args)
def filename(self, idx: int) -> str:
"""Override: return the image file name in the subset"""
return self._filenames[self._idx[self.pool[idx]]]
def image_size(self, idx: int) -> Tuple[int, int]:
"""Override: return the size (width, height) of an image in the subset"""
return self._image_sizes[self._idx[self.pool[idx]]]
@property
def anno_interaction(self) -> List[int]:
"""Override: Number of annotated box pairs for each interaction class"""
num_anno = [0 for _ in range(self.num_interation_cls)]
intra_idx = [self._idx[i] for i in self.pool]
for idx in intra_idx:
for hoi in self._anno[idx]['hoi']:
num_anno[hoi] += 1
return num_anno
@property
def anno_object(self) -> List[int]:
"""Override: Number of annotated box pairs for each object class"""
num_anno = [0 for _ in range(self.num_object_cls)]
anno_interaction = self.anno_interaction
for corr in self._class_corr:
num_anno[corr[1]] += anno_interaction[corr[0]]
return num_anno
@property
def anno_action(self) -> List[int]:
"""Override: Number of annotated box pairs for each action class"""
num_anno = [0 for _ in range(self.num_action_cls)]
anno_interaction = self.anno_interaction
for corr in self._class_corr:
num_anno[corr[2]] += anno_interaction[corr[0]]
return num_anno
class HICODet(ImageDataset):
"""
Arguments:
root(str): Root directory where images are downloaded to
anno_file(str): Path to json annotation file
transform(callable, optional): A function/transform that takes in an PIL image
and returns a transformed version
target_transform(callable, optional): A function/transform that takes in the
target and transforms it
transforms (callable, optional): A function/transform that takes input sample
and its target as entry and returns a transformed version.
"""
def __init__(self, root: str, anno_file: str,
transform: Optional[Callable] = None,
target_transform: Optional[Callable] = None,
transforms: Optional[Callable] = None) -> None:
super(HICODet, self).__init__(root, transform, target_transform, transforms)
with open(anno_file, 'r') as f:
anno = json.load(f)
import pdb;pdb.set_trace()
self.num_object_cls = 80
self.num_interation_cls = 600
self.num_action_cls = 117
self._anno_file = anno_file
# Load annotations
self._load_annotation_and_metadata(anno)
def __len__(self) -> int:
"""Return the number of images"""
return len(self._idx)
def __getitem__(self, i: int) -> tuple:
"""
Arguments:
i(int): Index to an image
Returns:
tuple[image, target]: By default, the tuple consists of a PIL image and a
dict with the following keys:
"boxes_h": list[list[4]]
"boxes_o": list[list[4]]
"hoi":: list[N]
"verb": list[N]
"object": list[N]
"""
intra_idx = self._idx[i]
return self._transforms(
self.load_image(os.path.join(self._root, self._filenames[intra_idx])),
self._anno[intra_idx]
)
def __repr__(self) -> str:
"""Return the executable string representation"""
reprstr = self.__class__.__name__ + '(root=' + repr(self._root)
reprstr += ', anno_file='
reprstr += repr(self._anno_file)
reprstr += ')'
# Ignore the optional arguments
return reprstr
def __str__(self) -> str:
"""Return the readable string representation"""
reprstr = 'Dataset: ' + self.__class__.__name__ + '\n'
reprstr += '\tNumber of images: {}\n'.format(self.__len__())
reprstr += '\tImage directory: {}\n'.format(self._root)
reprstr += '\tAnnotation file: {}\n'.format(self._root)
return reprstr
@property
def annotations(self) -> List[dict]:
return self._anno
@property
def class_corr(self) -> List[Tuple[int, int, int]]:
"""
Class correspondence matrix in zero-based index
[
[hoi_idx, obj_idx, verb_idx],
...
]
Returns:
list[list[3]]
"""
return self._class_corr.copy()
@property
def object_n_verb_to_interaction(self) -> List[list]:
"""
The interaction classes corresponding to an object-verb pair
HICODet.object_n_verb_to_interaction[obj_idx][verb_idx] gives interaction class
index if the pair is valid, None otherwise
Returns:
list[list[117]]
"""
lut = np.full([self.num_object_cls, self.num_action_cls], None)
for i, j, k in self._class_corr:
lut[j, k] = i
return lut.tolist()
@property
def object_to_interaction(self) -> List[list]:
"""
The interaction classes that involve each object type
Returns:
list[list]
"""
obj_to_int = [[] for _ in range(self.num_object_cls)]
for corr in self._class_corr:
obj_to_int[corr[1]].append(corr[0])
return obj_to_int
@property
def object_to_verb(self) -> List[list]:
"""
The valid verbs for each object type
Returns:
list[list]
"""
obj_to_verb = [[] for _ in range(self.num_object_cls)]
for corr in self._class_corr:
obj_to_verb[corr[1]].append(corr[2])
return obj_to_verb
@property
def anno_interaction(self) -> List[int]:
"""
Number of annotated box pairs for each interaction class
Returns:
list[600]
"""
return self._num_anno.copy()
@property
def anno_object(self) -> List[int]:
"""
Number of annotated box pairs for each object class
Returns:
list[80]
"""
num_anno = [0 for _ in range(self.num_object_cls)]
for corr in self._class_corr:
num_anno[corr[1]] += self._num_anno[corr[0]]
return num_anno
@property
def anno_action(self) -> List[int]:
"""
Number of annotated box pairs for each action class
Returns:
list[117]
"""
num_anno = [0 for _ in range(self.num_action_cls)]
for corr in self._class_corr:
num_anno[corr[2]] += self._num_anno[corr[0]]
return num_anno
@property
def objects(self) -> List[str]:
"""
Object names
Returns:
list[str]
"""
return self._objects.copy()
@property
def verbs(self) -> List[str]:
"""
Verb (action) names
Returns:
list[str]
"""
return self._verbs.copy()
@property
def interactions(self) -> List[str]:
"""
Combination of verbs and objects
Returns:
list[str]
"""
return [self._verbs[j] + ' ' + self.objects[i]
for _, i, j in self._class_corr]
def split(self, ratio: float) -> Tuple[HICODetSubset, HICODetSubset]:
"""
Split the dataset according to given ratio
Arguments:
ratio(float): The percentage of training set between 0 and 1
Returns:
train(Dataset)
val(Dataset)
"""
perm = np.random.permutation(len(self._idx))
n = int(len(perm) * ratio)
return HICODetSubset(self, perm[:n]), HICODetSubset(self, perm[n:])
def filename(self, idx: int) -> str:
"""Return the image file name given the index"""
return self._filenames[self._idx[idx]]
def image_size(self, idx: int) -> Tuple[int, int]:
"""Return the size (width, height) of an image"""
return self._image_sizes[self._idx[idx]]
def _load_annotation_and_metadata(self, f: dict) -> None:
"""
Arguments:
f(dict): Dictionary loaded from {anno_file}.json
"""
idx = list(range(len(f['filenames'])))
for empty_idx in f['empty']:
idx.remove(empty_idx)
num_anno = [0 for _ in range(self.num_interation_cls)]
for anno in f['annotation']:
for hoi in anno['hoi']:
num_anno[hoi] += 1
self._idx = idx
self._num_anno = num_anno
self._anno = f['annotation']
self._filenames = f['filenames']
self._image_sizes = f['size']
self._class_corr = f['correspondence']
self._empty_idx = f['empty']
self._objects = f['objects']
self._verbs = f['verbs'] |