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 bycreate_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 filebibs.csv
.
models
contains code fromhttps://github.com/tensorflow/models
. More specifically, it contains theobject_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
- If required on your environment, activate Tensor Flow:
source tensorflow/bin/activate
. - Set up the model:
bash run/setup_model.sh
. - Start the training job:
bash run/train.sh
. - Start the evaluation job:
bash run/eval.sh
.
Training the number recognizer
TODO