File size: 12,633 Bytes
a166479 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 |
"""
This interface provides access to four datasets:
1) refclef
2) refcoco
3) refcoco+
4) refcocog
split by unc and google
The following API functions are defined:
REFER - REFER api class
getRefIds - get ref ids that satisfy given filter conditions.
getAnnIds - get ann ids that satisfy given filter conditions.
getImgIds - get image ids that satisfy given filter conditions.
getCatIds - get category ids that satisfy given filter conditions.
loadRefs - load refs with the specified ref ids.
loadAnns - load anns with the specified ann ids.
loadImgs - load images with the specified image ids.
loadCats - load category names with the specified category ids.
getRefBox - get ref's bounding box [x, y, w, h] given the ref_id
showRef - show image, segmentation or box of the referred object with the ref
getMask - get mask and area of the referred object given ref
showMask - show mask of the referred object given ref
"""
import sys
import os.path as osp
import json
import pickle as pickle
import time
import itertools
import skimage.io as io
import matplotlib.pyplot as plt
from matplotlib.collections import PatchCollection
from matplotlib.patches import Polygon, Rectangle
from pprint import pprint
import numpy as np
from pycocotools import mask
class REFER:
def __init__(self, data_root, dataset='refcoco', splitBy='unc'):
# provide data_root folder which contains refclef, refcoco, refcoco+ and refcocog
# also provide dataset name and splitBy information
# e.g., dataset = 'refcoco', splitBy = 'unc'
print('loading dataset %s into memory...' % dataset)
if dataset == 'refcocog':
print('Split by {}!'.format(splitBy))
self.ROOT_DIR = osp.abspath(osp.dirname(__file__))
self.DATA_DIR = osp.join(data_root, dataset)
if dataset in ['refcoco', 'refcoco+', 'refcocog']:
self.IMAGE_DIR = osp.join(data_root, 'images/mscoco/images/train2014')
elif dataset == 'refclef':
self.IMAGE_DIR = osp.join(data_root, 'images/saiapr_tc-12')
else:
print('No refer dataset is called [%s]' % dataset)
sys.exit()
# load refs from data/dataset/refs(dataset).json
tic = time.time()
ref_file = osp.join(self.DATA_DIR, 'refs(' + splitBy + ').p')
self.data = {}
self.data['dataset'] = dataset
f = open(ref_file, 'r')
self.data['refs'] = pickle.load(open(ref_file, 'rb'))
# load annotations from data/dataset/instances.json
instances_file = osp.join(self.DATA_DIR, 'instances.json')
instances = json.load(open(instances_file, 'r'))
self.data['images'] = instances['images']
self.data['annotations'] = instances['annotations']
self.data['categories'] = instances['categories']
# create index
self.createIndex()
print('DONE (t=%.2fs)' % (time.time() - tic))
def createIndex(self):
# create sets of mapping
# 1) Refs: {ref_id: ref}
# 2) Anns: {ann_id: ann}
# 3) Imgs: {image_id: image}
# 4) Cats: {category_id: category_name}
# 5) Sents: {sent_id: sent}
# 6) imgToRefs: {image_id: refs}
# 7) imgToAnns: {image_id: anns}
# 8) refToAnn: {ref_id: ann}
# 9) annToRef: {ann_id: ref}
# 10) catToRefs: {category_id: refs}
# 11) sentToRef: {sent_id: ref}
# 12) sentToTokens: {sent_id: tokens}
print('creating index...')
# fetch info from instances
Anns, Imgs, Cats, imgToAnns = {}, {}, {}, {}
for ann in self.data['annotations']:
Anns[ann['id']] = ann
imgToAnns[ann['image_id']] = imgToAnns.get(ann['image_id'], []) + [ann]
for img in self.data['images']:
Imgs[img['id']] = img
for cat in self.data['categories']:
Cats[cat['id']] = cat['name']
# fetch info from refs
Refs, imgToRefs, refToAnn, annToRef, catToRefs = {}, {}, {}, {}, {}
Sents, sentToRef, sentToTokens = {}, {}, {}
for ref in self.data['refs']:
# ids
ref_id = ref['ref_id']
ann_id = ref['ann_id']
category_id = ref['category_id']
image_id = ref['image_id']
# add mapping related to ref
Refs[ref_id] = ref
imgToRefs[image_id] = imgToRefs.get(image_id, []) + [ref]
catToRefs[category_id] = catToRefs.get(category_id, []) + [ref]
refToAnn[ref_id] = Anns[ann_id]
annToRef[ann_id] = ref
# add mapping of sent
for sent in ref['sentences']:
Sents[sent['sent_id']] = sent
sentToRef[sent['sent_id']] = ref
sentToTokens[sent['sent_id']] = sent['tokens']
# create class members
self.Refs = Refs
self.Anns = Anns
self.Imgs = Imgs
self.Cats = Cats
self.Sents = Sents
self.imgToRefs = imgToRefs
self.imgToAnns = imgToAnns
self.refToAnn = refToAnn
self.annToRef = annToRef
self.catToRefs = catToRefs
self.sentToRef = sentToRef
self.sentToTokens = sentToTokens
print('index created.')
def getRefIds(self, image_ids=[], cat_ids=[], ref_ids=[], split=''):
image_ids = image_ids if type(image_ids) == list else [image_ids]
cat_ids = cat_ids if type(cat_ids) == list else [cat_ids]
ref_ids = ref_ids if type(ref_ids) == list else [ref_ids]
if len(image_ids) == len(cat_ids) == len(ref_ids) == len(split) == 0:
refs = self.data['refs']
else:
if not len(image_ids) == 0:
refs = [self.imgToRefs[image_id] for image_id in image_ids]
else:
refs = self.data['refs']
if not len(cat_ids) == 0:
refs = [ref for ref in refs if ref['category_id'] in cat_ids]
if not len(ref_ids) == 0:
refs = [ref for ref in refs if ref['ref_id'] in ref_ids]
if not len(split) == 0:
if split in ['testA', 'testB', 'testC']:
refs = [ref for ref in refs if split[-1] in ref['split']] # we also consider testAB, testBC, ...
elif split in ['testAB', 'testBC', 'testAC']:
refs = [ref for ref in refs if ref['split'] == split] # rarely used I guess...
elif split == 'test':
refs = [ref for ref in refs if 'test' in ref['split']]
elif split == 'train' or split == 'val':
refs = [ref for ref in refs if ref['split'] == split]
else:
print('No such split [%s]' % split)
sys.exit()
ref_ids = [ref['ref_id'] for ref in refs]
return ref_ids
def getAnnIds(self, image_ids=[], cat_ids=[], ref_ids=[]):
image_ids = image_ids if type(image_ids) == list else [image_ids]
cat_ids = cat_ids if type(cat_ids) == list else [cat_ids]
ref_ids = ref_ids if type(ref_ids) == list else [ref_ids]
if len(image_ids) == len(cat_ids) == len(ref_ids) == 0:
ann_ids = [ann['id'] for ann in self.data['annotations']]
else:
if not len(image_ids) == 0:
lists = [self.imgToAnns[image_id] for image_id in image_ids if
image_id in self.imgToAnns] # list of [anns]
anns = list(itertools.chain.from_iterable(lists))
else:
anns = self.data['annotations']
if not len(cat_ids) == 0:
anns = [ann for ann in anns if ann['category_id'] in cat_ids]
ann_ids = [ann['id'] for ann in anns]
if not len(ref_ids) == 0:
ids = set(ann_ids).intersection(set([self.Refs[ref_id]['ann_id'] for ref_id in ref_ids]))
return ann_ids
def getImgIds(self, ref_ids=[]):
ref_ids = ref_ids if type(ref_ids) == list else [ref_ids]
if not len(ref_ids) == 0:
image_ids = list(set([self.Refs[ref_id]['image_id'] for ref_id in ref_ids]))
else:
image_ids = self.Imgs.keys()
return image_ids
def getCatIds(self):
return self.Cats.keys()
def loadRefs(self, ref_ids=[]):
if type(ref_ids) == list:
return [self.Refs[ref_id] for ref_id in ref_ids]
elif type(ref_ids) == int:
return [self.Refs[ref_ids]]
def loadAnns(self, ann_ids=[]):
if type(ann_ids) == list:
return [self.Anns[ann_id] for ann_id in ann_ids]
elif type(ann_ids) == int or type(ann_ids) == unicode:
return [self.Anns[ann_ids]]
def loadImgs(self, image_ids=[]):
if type(image_ids) == list:
return [self.Imgs[image_id] for image_id in image_ids]
elif type(image_ids) == int:
return [self.Imgs[image_ids]]
def loadCats(self, cat_ids=[]):
if type(cat_ids) == list:
return [self.Cats[cat_id] for cat_id in cat_ids]
elif type(cat_ids) == int:
return [self.Cats[cat_ids]]
def getRefBox(self, ref_id):
ref = self.Refs[ref_id]
ann = self.refToAnn[ref_id]
return ann['bbox'] # [x, y, w, h]
def showRef(self, ref, seg_box='seg'):
ax = plt.gca()
# show image
image = self.Imgs[ref['image_id']]
I = io.imread(osp.join(self.IMAGE_DIR, image['file_name']))
ax.imshow(I)
# show refer expression
for sid, sent in enumerate(ref['sentences']):
print('%s. %s' % (sid + 1, sent['sent']))
# show segmentations
if seg_box == 'seg':
ann_id = ref['ann_id']
ann = self.Anns[ann_id]
polygons = []
color = []
c = 'none'
if type(ann['segmentation'][0]) == list:
# polygon used for refcoco*
for seg in ann['segmentation']:
poly = np.array(seg).reshape((len(seg) / 2, 2))
polygons.append(Polygon(poly, True, alpha=0.4))
color.append(c)
p = PatchCollection(polygons, facecolors=color, edgecolors=(1, 1, 0, 0), linewidths=3, alpha=1)
ax.add_collection(p) # thick yellow polygon
p = PatchCollection(polygons, facecolors=color, edgecolors=(1, 0, 0, 0), linewidths=1, alpha=1)
ax.add_collection(p) # thin red polygon
else:
# mask used for refclef
rle = ann['segmentation']
m = mask.decode(rle)
img = np.ones((m.shape[0], m.shape[1], 3))
color_mask = np.array([2.0, 166.0, 101.0]) / 255
for i in range(3):
img[:, :, i] = color_mask[i]
ax.imshow(np.dstack((img, m * 0.5)))
# show bounding-box
elif seg_box == 'box':
ann_id = ref['ann_id']
ann = self.Anns[ann_id]
bbox = self.getRefBox(ref['ref_id'])
box_plot = Rectangle((bbox[0], bbox[1]), bbox[2], bbox[3], fill=False, edgecolor='green', linewidth=3)
ax.add_patch(box_plot)
def getMask(self, ref):
# return mask, area and mask-center
ann = self.refToAnn[ref['ref_id']]
image = self.Imgs[ref['image_id']]
if type(ann['segmentation'][0]) == list: # polygon
rle = mask.frPyObjects(ann['segmentation'], image['height'], image['width'])
else:
rle = ann['segmentation']
m = mask.decode(rle)
m = np.sum(m, axis=2) # sometimes there are multiple binary map (corresponding to multiple segs)
m = m.astype(np.uint8) # convert to np.uint8
# compute area
area = sum(mask.area(rle)) # should be close to ann['area']
return {'mask': m, 'area': area}
def showMask(self, ref):
M = self.getMask(ref)
msk = M['mask']
ax = plt.gca()
ax.imshow(msk)
if __name__ == '__main__':
refer = REFER(dataset='refcocog', splitBy='google')
ref_ids = refer.getRefIds()
ref_ids = refer.getRefIds(split='train')
print('There are %s training referred objects.' % len(ref_ids))
for ref_id in ref_ids:
ref = refer.loadRefs(ref_id)[0]
if len(ref['sentences']) < 2:
continue
print('The label is %s.' % refer.Cats[ref['category_id']])
plt.figure()
refer.showRef(ref, seg_box='box')
plt.show()
|