Wider Face Dataset to TF Record

Convert the WIDER Face to TFRecord format.

Important: this project runs under Python3.

Why doing this?

The TFRecord format will allow you to use the Tensorflow Object Detection API in order to train a new model based on Transfer Learning from many available pretrained models.

How to run

1. Clone the repo
2. [Optional] Create a virtual environment to keep dependencies isolated.
3. Run pip install -r requirements.txt
4. Configure config.py file with your appropiate paths.
   • You should have downloaded from here the dataset and the annotations' folder
   • If you don't want a set to be converted then leave the field in None. For example, if test is not going to be converted you should leave TEST_WIDER_PATH in None.
5. Run python wider_to_tfrecord.py
6. When done, tfrecord files should be in the output folder you specified in config.py.

About testing images set

Testing images do not contain ground truth bounding boxes. In case you want to convert the testing set as well, the test.tfrecord file will contain images only without bounding boxes.

Acknowledgement

• This work was based on this original version for Python2.
• The conversion follows the standards indicated in the official Tensorflow Object Detection API site here