We're trying to train a bib detector here.

Directory structure

• data_raw is just the place where the images are downloaded and wait to be annotated.
• data has annotated data.
  • data/bib_detection/annotations has a collection of images in the format given by VGG Image Annotator. These images are ready to be processed by create_tf_records.py to build the TF examples.
  • data/number_recognition/annotations contains images of bibs. Some of the images were cropped using the data from ground truth for bib detection and others were cropped using the predictions of the bib detector. All annotations are kept in a single file bibs.csv.
• models contains code from https://github.com/tensorflow/models. More specifically, it contains the object_detection & attention_ocr models.

Training the bib detector

Annotations

The images are grouped in directories by event. Each directory contains a few images and one .csv file which contains the annotations.

This structure needs to be kept strictly or the create_tf_records.py script won't work correctly.

How to start training

1. If required on your environment, activate Tensor Flow: source tensorflow/bin/activate.
2. Set up the model: bash run/setup_model.sh.
3. Start the training job: bash run/train.sh.
4. Start the evaluation job: bash run/eval.sh.

Training the number recognizer

TODO