def __init__(self,
model_dir,
in_dir,
out_dir,
nms_th_for_all_scale=0.5,
score_th=0.2,
scales=([384, 384], [768, 384], [768, 768]),
min_side_scale=384,
save_res_path='eval_res.txt'
):
if os.path.exists(in_dir):
self.in_dir = in_dir
else:
raise ValueError('{} does not existed!!!'.format(in_dir))
self.out_dir = out_dir
self.suffixes = ['.png', '.PNG', '.jpg', '.jpeg']
self.img_path, self.img_num = self.get_img_path()
self.nms_th_for_all_scale = nms_th_for_all_scale
self.nms_threshold = 0.45
self.score_th = score_th
print('self.score_th', self.score_th)
self.make_out_dir()
self.text_scales = scales
self.data_format = 'NHWC'
self.select_threshold = 0.01
self.min_side_scale = min_side_scale
self.max_side_scale = self.min_side_scale * 2 # 384 * 2
self.save_xml_flag = True
self.save_txt_flag = True
self.dynamic_scale_flag = False
self.allow_padding = False
self.allow_post_processing = False
self.allow_eval_flag = False
self.resize_flag = False
self.save_eval_resut_path = save_res_path
self.model_path = None
self.config = tf.ConfigProto(allow_soft_placement=True)
self.config.gpu_options.allow_growth = True
self.graph = tf.Graph()
self.session_text = tf.Session(graph=self.graph, config=self.config)
with self.session_text.as_default():
with self.graph.as_default():
self.img_text = tf.placeholder(
tf.float32, shape=(None, None, 3))
print(len(self.text_scales))
self.scale_text = tf.placeholder(tf.int32, shape=(2))
img_pre_text, label_pre_text, bboxes_pre_text, self.bboxes_img_text, xs_text, ys_text = ssd_vgg_preprocessing.preprocess_for_eval(
self.img_text,
None,
None,
None,
None,
self.scale_text,
self.data_format,
resize=ssd_vgg_preprocessing.Resize.WARP_RESIZE)
image_text_4d = tf.expand_dims(img_pre_text, 0)
image_text_4d = tf.cast(image_text_4d, tf.float32)
self.image_text_4d = image_text_4d
self.net_text = txtbox_384.TextboxNet()
with slim.arg_scope(
self.net_text.arg_scope(data_format=self.data_format)):
self.predictions_text, self.localisations_text, self.logits_text, self.endpoints_text, self.l_shape = self.net_text.net(
self.image_text_4d,
is_training=False,
reuse=tf.AUTO_REUSE,
update_feat_shapes=True)
saver_text = tf.train.Saver()
if os.path.isdir(model_dir):
ckpt_path = tf.train.latest_checkpoint(model_dir)
self.model_path = os.path.join(model_dir, ckpt_path)
else:
ckpt_path = model_dir
self.model_path = ckpt_path
print(model_dir)
saver_text.restore(self.session_text, ckpt_path)
logging.info("Textbox++ model initialized.")
config = tf.compat.v1.ConfigProto(log_device_placement=False,
gpu_options=gpu_options)
isess = tf.compat.v1.Session(config=config)
# Input placeholder.
net_shape = (384, 384)
#net_shape = (768, 768)
data_format = 'NHWC'
img_input = tf.compat.v1.placeholder(tf.float32, shape=(None, None, 3))
# Evaluation pre-processing: resize to SSD net shape.
image_pre, labels_pre, bboxes_pre, bbox_img, xs, ys = ssd_vgg_preprocessing.preprocess_for_eval(
img_input,
None,
None,
None,
None,
net_shape,
data_format,
resize=ssd_vgg_preprocessing.Resize.WARP_RESIZE)
image_4d = tf.expand_dims(image_pre, 0)
image_4d = tf.cast(image_4d, tf.float32)
# Define the txt_box model.
reuse = True if 'txt_net' in locals() else None
txt_net = txtbox_384.TextboxNet()
print(txt_net.params.img_shape)
print('reuse:', reuse)
with slim.arg_scope(txt_net.arg_scope(data_format=data_format)):
predictions, localisations, logits, end_points = txt_net.net(