def array_to_mtcnndatum(img_data, label, roi, pts=None):
"""Convert data to mtcnnDatum
"""
mtcnn_datum = caffe_pb2.MTCNNDatum()
datum = caffe_pb2.Datum()
datum = caffe.io.array_to_datum(img_data.transpose((2, 0, 1)),
label) # rgb
# datum = caffe.io.array_to_datum(img_data, label) #bgr
# copy data
mtcnn_datum.datum.data = datum.data
mtcnn_datum.datum.label = datum.label
mtcnn_datum.datum.channels = datum.channels
mtcnn_datum.datum.height = datum.height
mtcnn_datum.datum.width = datum.width
# set roi data
mtcnn_datum.rois.xmin = float(roi[0])
mtcnn_datum.rois.ymin = float(roi[1])
mtcnn_datum.rois.xmax = float(roi[2])
mtcnn_datum.rois.ymax = float(roi[3])
if pts is not None:
mtcnn_datum.pts = pts
return mtcnn_datum
num_for_each = 1
# create the lmdb file
# map_size指的是数据库的最大容量，根据需求设置
if (lmdb_id == 0):
lmdb_env_12 = lmdb.open(dir_prefix + 'mtcnn_train_12', map_size=1000000000)
lmdb_txn_12 = lmdb_env_12.begin(write=True)
elif (lmdb_id == 1):
lmdb_env_24 = lmdb.open(dir_prefix + 'mtcnn_train_24', map_size=5000000000)
lmdb_txn_24 = lmdb_env_24.begin(write=True)
else:
lmdb_env_48 = lmdb.open(dir_prefix + 'mtcnn_train_48',
map_size=10000000000)
lmdb_txn_48 = lmdb_env_48.begin(write=True)
# 因为caffe中经常采用datum这种数据结构存储数据
mtcnn_datum = caffe_pb2.MTCNNDatum()
for line_idx, annotation in enumerate(annotations):
annotation = annotation.strip().split(' ') #每一行的数据以空白分隔符为界限
im_path = annotation[0] #图片的路径
bbox = map(float, annotation[1:])
if np.size(bbox) % 4 != 0: #标注数据有问题
print "the annotation data in line %d is invalid, please check file %s !" % (
line_idx + 1, anno_file)
exit(-1)
elif np.size(bbox) == 0:
continue
boxes = np.array(bbox, dtype=np.float32).reshape(-1, 4)