An implementation of Tfrecord in muti-oriented detection task when data augmentation is needed. The bounding box annotation here is x1,y1,x2,y2,x3,y3,x4,y4 which is often used in text detection.

• Python2.7
• TensorFlow1.0

One image corresponds to a txt file. The annotations format in txt file is as follows:

x1,y1,x2,y2,x3,y3,x4,y4,label_name
x1,y1,x2,y2,x3,y3,x4,y4,label_name

You should prepare both train and val datasets. The file struture should be like this:

-$ROOT_PATH
  -Dataset_train
    -JPEGImages
      -your images
    -Annotations
      -your txt_file	
  -Dataset_val
    -JPEGImages
      -your images
    -Annotations
      -your txt_file

1. First, you should set the right path in config.py and the label names in label_dict.py.

2. Create .tfrecord files.

python txt_to_tfrecord.py --type train
python txt_to_tfrecord.py --type val

3. When training, use the function next_batch in read_tfrecord.py.