conv_pascal_to_csv

Annotation format converter from Pascal VOC format to fizyr/keras-retinanet csv format

This script improve object detection development workflow with fizyr/keras-retinanet.

Description

Around object detection development, it is a hard work to preparing training data. Some software (ex, LabelImg) help this situation but doesn't output with csv format of fizyr/keras-retinanet.

Using this script conv_pascal_to_csv.rb, you can convert annotations from Pascal VOC xml format with which LabelImg outputs, to csv format which fizyr/keras-retinanet requires.

Requirement

Ruby 2 (tested with ruby 2.5.1)

Demo

git clone https://github.com/akchan/conv_pascal_to_csv.git
cd conv_pascal_to_csv/sample/retinanet_csv_format
ruby ../../conv_pascal_to_csv.rb --annotation-path ../PascalVOC_format/Annotations --image-path ../PascalVOC_format/JPEGImages

Then, conv_pascal_to_csv.rb creates some files/directories like below.

retinanet_csv_format
- csv
  - annotations.csv
  - val_annotations.csv
  - classes.csv
- images
  - jpeg files

You can give these directory to train.py of fizyr/keras-retinanet like below.

cd PATH_TO_KERAS_RETINANET
python ./keras_retinanet/bin/train.py csv PATH_TO_ANNOTATIONS PATH_TO_CLASSES --val-annotations PATH_TO_VAL_ANNOTATIONS

Sources

Sample images

Licence

MIT licence

See LICENCE file.

Author

akchan (Satoshi Funayama, GitHub, twitter)