def _load_trainvalsplit(self, ):
assert self._split in ['trainval2014']
annotation_dir = os.path.join(self._data_dir, 'annotations')
minival_path = os.path.join(annotation_dir,
'instances_minival2014.json')
minival2014_url = 'https://dl.dropboxusercontent.com/s/o43o90bna78omob/instances_minival2014.json.zip?dl=0'
assert os.path.exists(
minival_path), 'need to download %s minival split to %s' % (
minival2014_url, minival_path)
import ujson as json
with open(minival_path, 'r') as f:
self._minival = json.load(f)
annFile = os.path.join(annotation_dir, 'instances_train2014.json')
coco_train = COCO(annFile)
annFile = os.path.join(annotation_dir, 'instances_val2014.json')
coco_val = COCO(annFile)
imgs1 = [(img_id, coco_train.imgs[img_id], 0)
for img_id in coco_train.imgs if img_id != 320612]
imgs2 = [(img_id, coco_val.imgs[img_id], 1) for img_id in coco_val.imgs
if not self._is_in_minival(img_id) and img_id != 320612]
imgs = imgs1 + imgs2
self._data_size = len(imgs)
self._imgs = imgs
self._cocos = (coco_train, coco_val)
self.classes = [u'background'] + [
cls['name'] for cls in coco_train.loadCats(coco_train.getCatIds())
]
return
def _load_trainvalsplit(self, ):
assert self._split in ['trainval2014']
annotation_dir = os.path.join(self._data_dir, 'annotations')
annFile = os.path.join(annotation_dir, 'instances_trainval2014.json')
coco_train = COCO(annFile)
imgs = [(img_id, coco_train.imgs[img_id], 0)
for img_id in coco_train.imgs if img_id != 320612]
self._data_size = len(imgs)
self._imgs = imgs
self._cocos = (coco_train, )
self.classes = [u'background'] + [
cls['name'] for cls in coco_train.loadCats(coco_train.getCatIds())
]
return
def _load(self, ):
assert self._split in ['train2014', 'val2014', 'minival2014']
annFile = os.path.join(self._data_dir, 'annotations',
'instances_%s.json' % (self._split))
coco = COCO(annFile)
# imgs = [(img_id, coco.imgs[img_id]) for img_id in coco.imgs]
imgs = []
for img_id in coco.imgs:
if img_id == 320612:
continue
imgs.append((img_id, coco.imgs[img_id], 0))
self._data_size = len(imgs)
self._imgs = imgs
self._cocos = (coco, )
self.classes = [u'background'] + [
cls['name'] for cls in coco.loadCats(coco.getCatIds())
]
return
def _add_to_tfrecord_trainvalsplit(record_dir, image_dir, annotation_dir,
split_name):
"""Loads image files and writes files to a TFRecord.
Note: masks and bboxes will lose shape info after converting to string.
"""
assert split_name in ['trainval2014', 'minival2014']
# NOTE: this instances_minival2014.json file cannot be processed by official COCO API,
# so just use its id list, ['images']['id']
minival_path = os.path.join(annotation_dir, 'instances_minival2014.json')
minival2014_url = 'https://dl.dropboxusercontent.com/s/o43o90bna78omob/instances_minival2014.json.zip?dl=0'
assert os.path.exists(
minival_path), 'need to download %s minival split to %s' % (
minival2014_url, minival_path)
import ujson as json
with open(minival_path, 'r') as f:
minival = json.load(f)
def is_in_minival(img_id, minival):
for img in minival['images']:
if (img['id']) == (img_id):
return True
return False
annFile = os.path.join(annotation_dir, 'instances_train2014.json')
coco_train = COCO(annFile)
annFile = os.path.join(annotation_dir, 'instances_val2014.json')
coco_val = COCO(annFile)
cats = coco_train.loadCats(coco_train.getCatIds())
# imgs = [(img_id, coco_train.imgs[img_id]) for img_id in coco_train.imgs] + \
# [(img_id, coco_val.imgs[img_id]) for img_id in coco_val.imgs]
imgs1 = [(img_id, coco_train.imgs[img_id]) for img_id in coco_train.imgs]
imgs2 = [(img_id, coco_val.imgs[img_id]) for img_id in coco_val.imgs]
imgs = imgs1 + imgs2
num_of_train = len(coco_train.imgs)
num_of_all = len(imgs)
num_per_shard = 2500
num_shards = int(
np.ceil((len(imgs) + 0.0 - len(minival['images'])) / num_per_shard))
if split_name == 'minival2014':
num_shards = int(
np.ceil((len(minival['images']) + 0.0) / num_per_shard))
with tf.Graph().as_default(), tf.device('/cpu:0'):
image_reader = ImageReader()
# encode mask to png_string
mask_placeholder = tf.placeholder(dtype=tf.uint8)
encoded_image = tf.image.encode_png(mask_placeholder)
with tf.Session('') as sess:
cnt = 0
shard_id = -1
for i in range(len(imgs)):
img_id = imgs[i][0]
img_name = imgs[i][1]['file_name']
split = img_name.split('_')[1]
img_name = os.path.join(image_dir, split, img_name)
if str(img_id) == '320612':
continue
is_minival = is_in_minival(img_id, minival)
if split_name == 'trainval2014' and is_minival:
continue
if split_name == 'minival2014' and not is_minival:
continue
cnt += 1
if cnt % num_per_shard == 1:
shard_id += 1
record_filename = _get_dataset_filename(
record_dir, split_name, shard_id, num_shards)
options = tf.python_io.TFRecordOptions(
TFRecordCompressionType.ZLIB)
tfrecord_writer = tf.python_io.TFRecordWriter(
record_filename, options=options)
if cnt % 100 == 1:
print('%d (image_id: %d) of %d, split: %s, shard_id: %d' %
(i, img_id, len(imgs), split_name, shard_id))
# process anns
height, width = imgs[i][1]['height'], imgs[i][1]['width']
coco = coco_train if i < num_of_train else coco_val
gt_boxes, masks, mask = _get_coco_masks(
coco, img_id, height, width, img_name)
# read image as RGB numpy
img = np.array(Image.open(img_name))
if img.size == height * width:
print('Gray Image %s' % str(img_id))
im = np.empty((height, width, 3), dtype=np.uint8)
im[:, :, :] = img[:, :, np.newaxis]
img = im
img = img.astype(np.uint8)
assert img.size == width * height * 3, '%s' % str(img_id)
img_raw = img.tostring()
mask_raw = mask.tostring()
example = _to_tfexample_coco_raw(img_id, img_raw, mask_raw,
height, width,
gt_boxes.shape[0],
gt_boxes.tostring(),
masks.tostring())
tfrecord_writer.write(example.SerializeToString())
if cnt % num_per_shard == 0 or i == len(imgs) - 1:
tfrecord_writer.close()
def _add_to_tfrecord(record_dir, image_dir, annotation_dir, split_name):
"""Loads image files and writes files to a TFRecord.
Note: masks and bboxes will lose shape info after converting to string.
"""
assert split_name in [
'train2014', 'val2014', 'valminusminival2014', 'minival2014'
]
annFile = os.path.join(annotation_dir, 'instances_%s.json' % (split_name))
coco = COCO(annFile)
cats = coco.loadCats(coco.getCatIds())
print('%s has %d images' % (split_name, len(coco.imgs)))
imgs = [(img_id, coco.imgs[img_id]) for img_id in coco.imgs]
num_shards = int(len(imgs) / 2500)
num_per_shard = int(math.ceil(len(imgs) / float(num_shards)))
with tf.Graph().as_default(), tf.device('/cpu:0'):
image_reader = ImageReader()
# encode mask to png_string
mask_placeholder = tf.placeholder(dtype=tf.uint8)
encoded_image = tf.image.encode_png(mask_placeholder)
with tf.Session('') as sess:
for shard_id in range(num_shards):
record_filename = _get_dataset_filename(
record_dir, split_name, shard_id, num_shards)
options = tf.python_io.TFRecordOptions(
TFRecordCompressionType.ZLIB)
with tf.python_io.TFRecordWriter(
record_filename, options=options) as tfrecord_writer:
start_ndx = shard_id * num_per_shard
end_ndx = min((shard_id + 1) * num_per_shard, len(imgs))
for i in range(start_ndx, end_ndx):
if i % 50 == 0:
sys.stdout.write(
'\r>> Converting image %d/%d shard %d\n' %
(i + 1, len(imgs), shard_id))
sys.stdout.flush()
# image id and path
img_id = imgs[i][0]
img_name = imgs[i][1]['file_name']
split = img_name.split('_')[1]
img_name = os.path.join(image_dir, split, img_name)
if FLAGS.vis:
im = Image.open(img_name)
im.save('img.png')
plt.figure(0)
plt.axis('off')
plt.imshow(im)
# plt.show()
# plt.close()
# jump over the damaged images
if str(img_id) == '320612':
continue
# process anns
height, width = imgs[i][1]['height'], imgs[i][1][
'width']
gt_boxes, masks, mask = _get_coco_masks(
coco, img_id, height, width, img_name)
# read image as RGB numpy
img = np.array(Image.open(img_name))
if img.size == height * width:
print('Gray Image %s' % str(img_id))
im = np.empty((height, width, 3), dtype=np.uint8)
im[:, :, :] = img[:, :, np.newaxis]
img = im
img = img.astype(np.uint8)
assert img.size == width * height * 3, '%s' % str(
img_id)
img_raw = img.tostring()
mask_raw = mask.tostring()
example = _to_tfexample_coco_raw(
img_id, img_raw, mask_raw, height, width,
gt_boxes.shape[0], gt_boxes.tostring(),
masks.tostring())
tfrecord_writer.write(example.SerializeToString())
sys.stdout.write('\n')
sys.stdout.flush()
def _add_to_tfrecord(record_dir, image_dir, annotation_dir, split_name):
"""Loads image files and writes files to a TFRecord.
Note: masks and bboxes will lose shape info after converting to string.
"""
assert split_name in ['train2014', 'val2014']
annFile = os.path.join(annotation_dir, 'instances_%s.json' % (split_name))
coco = COCO(annFile)
cats = coco.loadCats(coco.getCatIds())
print('%s has %d images' % (split_name, len(coco.imgs)))
imgs = [(img_id, coco.imgs[img_id]) for img_id in coco.imgs]
num_shards = 40 if split_name == 'train2014' else 20
num_per_shard = int(math.ceil(len(imgs) / float(num_shards)))
with tf.Graph().as_default(), tf.device('/cpu:0'):
image_reader = ImageReader()
# encode mask to png_string
mask_placeholder = tf.placeholder(dtype=tf.uint8)
encoded_image = tf.image.encode_png(mask_placeholder)
with tf.Session('') as sess:
for shard_id in range(num_shards):
record_filename = _get_dataset_filename(
record_dir, split_name, shard_id, num_shards)
options = tf.python_io.TFRecordOptions(
TFRecordCompressionType.ZLIB)
with tf.python_io.TFRecordWriter(
record_filename, options=options) as tfrecord_writer:
start_ndx = shard_id * num_per_shard
end_ndx = min((shard_id + 1) * num_per_shard, len(imgs))
for i in range(start_ndx, end_ndx):
sys.stdout.write(
'\r>> Converting image %d/%d shard %d\n' %
(i + 1, len(imgs), shard_id))
sys.stdout.flush()
# image id and path
img_id = imgs[i][0]
img_name = imgs[i][1]['file_name']
img_name = os.path.join(image_dir, split_name,
img_name)
if FLAGS.vis:
im = Image.open(img_name)
im.save('img.png')
plt.figure(0)
plt.axis('off')
plt.imshow(im)
# plt.show()
# plt.close()
# jump over the damaged images
if split_name == 'val2014' and str(img_id) == '320612':
continue
# process anns
h, w = imgs[i][1]['height'], imgs[i][1]['width']
classes, bboxes, masks, mask = _get_coco_masks(
coco, img_id, h, w)
assert classes.shape[0] == bboxes.shape[0] == masks.shape[0], \
'Check number of instances for %s' % (img_name)
# this encode matrix to png format string buff
label_data = sess.run(encoded_image,
feed_dict={
mask_placeholder:
np.expand_dims(mask, axis=2)
})
# read image
assert os.path.exists(
img_name), '%s dont exists' % img_name
image_data = tf.gfile.FastGFile(img_name, 'r').read()
height, width, depth = image_reader.read_jpeg_dims(
sess, image_data)
# to tf-record
example = _to_tfexample_v2(image_data, 'jpg',
label_data, 'png', height,
width, classes.shape[0],
classes.tolist(),
bboxes.tostring(),
masks.tostring())
tfrecord_writer.write(example.SerializeToString())
sys.stdout.write('\n')
sys.stdout.flush()
def _add_to_tfrecord_trainvalsplit(record_dir, image_dir, annotation_dir, split_name):
"""Loads image files and writes files to a TFRecord.
Note: masks and bboxes will lose shape info after converting to string.
"""
assert split_name in ['trainval2014', 'minival2014']
# NOTE: this instances_minival2014.json file cannot be processed by official COCO API,
# so just use its id list, ['images']['id']
minival_path = os.path.join(annotation_dir, 'instances_minival2014.json')
minival2014_url='https://dl.dropboxusercontent.com/s/o43o90bna78omob/instances_minival2014.json.zip?dl=0'
assert os.path.exists(minival_path), 'need to download %s minival split to %s' %(minival2014_url, minival_path)
import ujson as json
with open(minival_path, 'r') as f:
minival = json.load(f)
def is_in_minival(img_id, minival):
for img in minival['images']:
if (img['id']) == (img_id):
return True
return False
annFile = os.path.join(annotation_dir, 'instances_train2014.json')
coco_train = COCO(annFile)
annFile = os.path.join(annotation_dir, 'instances_val2014.json')
coco_val = COCO(annFile)
cats = coco_train.loadCats(coco_train.getCatIds())
# imgs = [(img_id, coco_train.imgs[img_id]) for img_id in coco_train.imgs] + \
# [(img_id, coco_val.imgs[img_id]) for img_id in coco_val.imgs]
imgs1 = [(img_id, coco_train.imgs[img_id]) for img_id in coco_train.imgs]
imgs2 = [(img_id, coco_val.imgs[img_id]) for img_id in coco_val.imgs]
imgs = imgs1 + imgs2
num_of_train = len(coco_train.imgs)
num_of_all = len(imgs)
num_per_shard = 2500
num_shards = int(np.ceil((len(imgs) + 0.0 - len(minival['images'])) / num_per_shard))
if split_name == 'minival2014':
num_shards = int(np.ceil((len(minival['images']) + 0.0) / num_per_shard))
with tf.Graph().as_default(), tf.device('/cpu:0'):
image_reader = ImageReader()
# encode mask to png_string
mask_placeholder = tf.placeholder(dtype=tf.uint8)
encoded_image = tf.image.encode_png(mask_placeholder)
with tf.Session('') as sess:
cnt = 0
shard_id = -1
for i in range(len(imgs)):
img_id = imgs[i][0]
img_name = imgs[i][1]['file_name']
split = img_name.split('_')[1]
img_name = os.path.join(image_dir, split, img_name)
if str(img_id) == '320612':
continue
is_minival = is_in_minival(img_id, minival)
if split_name == 'trainval2014' and is_minival:
continue
if split_name == 'minival2014' and not is_minival:
continue
cnt += 1
if cnt % num_per_shard == 1:
shard_id += 1
record_filename = _get_dataset_filename(record_dir, split_name, shard_id, num_shards)
options = tf.python_io.TFRecordOptions(TFRecordCompressionType.ZLIB)
tfrecord_writer = tf.python_io.TFRecordWriter(record_filename, options=options)
if cnt % 100 == 1:
print ('%d (image_id: %d) of %d, split: %s, shard_id: %d' %(i, img_id, len(imgs), split_name, shard_id))
# process anns
height, width = imgs[i][1]['height'], imgs[i][1]['width']
coco = coco_train if i < num_of_train else coco_val
gt_boxes, masks, mask = _get_coco_masks(coco, img_id, height, width, img_name)
# read image as RGB numpy
img = np.array(Image.open(img_name))
if img.size == height * width:
print ('Gray Image %s' % str(img_id))
im = np.empty((height, width, 3), dtype=np.uint8)
im[:, :, :] = img[:, :, np.newaxis]
img = im
img = img.astype(np.uint8)
assert img.size == width * height * 3, '%s' % str(img_id)
img_raw = img.tostring()
mask_raw = mask.tostring()
example = _to_tfexample_coco_raw(
img_id,
img_raw,
mask_raw,
height, width, gt_boxes.shape[0],
gt_boxes.tostring(), masks.tostring())
tfrecord_writer.write(example.SerializeToString())
if cnt % num_per_shard == 0 or i == len(imgs)-1:
tfrecord_writer.close()
def _add_to_tfrecord(record_dir, image_dir, annotation_dir, split_name):
"""Loads image files and writes files to a TFRecord.
Note: masks and bboxes will lose shape info after converting to string.
"""
assert split_name in ['train2014', 'val2014', 'valminusminival2014', 'minival2014']
annFile = os.path.join(annotation_dir, 'instances_%s.json' % (split_name))
coco = COCO(annFile)
cats = coco.loadCats(coco.getCatIds())
print ('%s has %d images' %(split_name, len(coco.imgs)))
imgs = [(img_id, coco.imgs[img_id]) for img_id in coco.imgs]
num_shards = int(len(imgs) / 2500)
num_per_shard = int(math.ceil(len(imgs) / float(num_shards)))
with tf.Graph().as_default(), tf.device('/cpu:0'):
image_reader = ImageReader()
# encode mask to png_string
mask_placeholder = tf.placeholder(dtype=tf.uint8)
encoded_image = tf.image.encode_png(mask_placeholder)
with tf.Session('') as sess:
for shard_id in range(num_shards):
record_filename = _get_dataset_filename(record_dir, split_name, shard_id, num_shards)
options = tf.python_io.TFRecordOptions(TFRecordCompressionType.ZLIB)
with tf.python_io.TFRecordWriter(record_filename, options=options) as tfrecord_writer:
start_ndx = shard_id * num_per_shard
end_ndx = min((shard_id + 1) * num_per_shard, len(imgs))
for i in range(start_ndx, end_ndx):
if i % 50 == 0:
sys.stdout.write('\r>> Converting image %d/%d shard %d\n' % (
i + 1, len(imgs), shard_id))
sys.stdout.flush()
# image id and path
img_id = imgs[i][0]
img_name = imgs[i][1]['file_name']
split = img_name.split('_')[1]
img_name = os.path.join(image_dir, split, img_name)
if FLAGS.vis:
im = Image.open(img_name)
im.save('img.png')
plt.figure(0)
plt.axis('off')
plt.imshow(im)
# plt.show()
# plt.close()
# jump over the damaged images
if str(img_id) == '320612':
continue
# process anns
height, width = imgs[i][1]['height'], imgs[i][1]['width']
gt_boxes, masks, mask = _get_coco_masks(coco, img_id, height, width, img_name)
# read image as RGB numpy
img = np.array(Image.open(img_name))
if img.size == height * width:
print ('Gray Image %s' % str(img_id))
im = np.empty((height, width, 3), dtype=np.uint8)
im[:, :, :] = img[:, :, np.newaxis]
img = im
img = img.astype(np.uint8)
assert img.size == width * height * 3, '%s' % str(img_id)
img_raw = img.tostring()
mask_raw = mask.tostring()
example = _to_tfexample_coco_raw(
img_id,
img_raw,
mask_raw,
height, width, gt_boxes.shape[0],
gt_boxes.tostring(), masks.tostring())
tfrecord_writer.write(example.SerializeToString())
sys.stdout.write('\n')
sys.stdout.flush()
import skimage.io as io
import pylab
import json
pylab.rcParams['figure.figsize'] = (10.0, 8.0)
annType = ['segm', 'bbox', 'keypoints']
annType = annType[0] #specify type here
prefix = 'person_keypoints' if annType == 'keypoints' else 'instances'
print 'Running demo for *%s* results.' % (annType)
#initialize COCO ground truth api
dataDir = 'data/coco/'
dataType = 'train2014' #val2014
annFile = '%s/annotations/%s_%s.json' % (dataDir, prefix, dataType)
cocoGt = COCO(annFile)
#initialize COCO detections api
# resFile='%s/results/%s_%s_fake%s100_results.json'
# resFile = resFile%(dataDir, prefix, dataType, annType)
resFile = 'output/mask_rcnn/results.json'
cocoDt = cocoGt.loadRes(resFile)
with open(resFile) as results:
res = json.load(results)
imgIds = []
for inst in res:
imgIds.append(inst['image_id'])
def mat2coco_annots(annot_path):
for split, annot_path in zip(['train', 'val'], [
'data/citypersons/annotations/anno_train.mat',
'data/citypersons/annotations/anno_val.mat'
]):
annots = sio.loadmat(annot_path)
annots = annots['anno_' + split + '_aligned'].reshape([-1])
coco = COCO()
coco.dataset['images'] = []
coco.dataset['categories'] = []
coco.dataset['annotations'] = []
coco.dataset['categories'].append({
'name': 'pedestrian',
'id': 1,
'supercategory': 'person',
})
coco.dataset['categories'].append({
'name': 'rider',
'id': 2,
'supercategory': 'person',
})
coco.dataset['categories'].append({
'name': 'sitting_person',
'id': 3,
'supercategory': 'person',
})
coco.dataset['categories'].append({
'name': 'other',
'id': 4,
'supercategory': 'person',
})
annid = 0
for i, ann in enumerate(annots):
ann = ann.reshape([-1])
city_name, image_name, bbs = ann[0][0][0], ann[0][1][0], ann[0][2]
coco.dataset['images'].append({
'id': i,
'file_name': image_name,
'height': 1024,
'width': 2048,
'url': 'citypersons',
})
gt_classes, gt_boxes = bbs[:, 0], bbs[:, 1:5]
areas = (gt_boxes[:, 2] + 1) * (gt_boxes[:, 3] + 1)
heights = gt_boxes[:, 3].copy()
gt_boxes[:, 2:4] += gt_boxes[:, 0:2]
for j in range(gt_classes.size):
bb = gt_boxes[j, :]
x1, y1, x2, y2 = [bb[0], bb[1], bb[2], bb[3]]
coco.dataset['annotations'].append({
'image_id':
i,
'bbox':
gt_boxes[j, :],
'category_id':
gt_classes[j],
'id':
annid + 1,
'iscrowd':
0,
'area':
areas[j],
'segmentation': [[x1, y1, x1, y2, x2, y2, x2, y1]],
'height':
heights[j],
})
coco.createIndex()
return coco
gt_classes[j],
'id':
annid + 1,
'iscrowd':
0,
'area':
areas[j],
'segmentation': [[x1, y1, x1, y2, x2, y2, x2, y1]],
'height':
heights[j],
})
coco.createIndex()
return coco
if __name__ == '__main__':
d = json.load(open(annot_path, 'r'))
print(d.keys())
print(d['annotations'][0].keys(), d['annotations'][0])
print(d['categories'][0].keys(), d['categories'][0], len(d['categories']))
print(d['images'][0], d['images'][0]['id'])
annot_path_trainval = './data/coco/annotations/instances_trainval_minus2014.json'
if not os.path.exists(annot_path_trainval):
merge_trainval(annot_path_trainval)
with open(annot_path_new, 'w') as f:
json.dump(d, f, indent=1, separators=(',', ': '))
train_annots = COCO(annot_path_trainval)