OpenImages_CustomDataset

A data preprocessing script for creating your own custom dataset out of OpenImagesDB

OpenImages is an extensive dataset made by google. I have arranged this simple script for those interested in assembling your own dataset, as a subset from OpenImages. As default, I use a subset called BodyParts (BP), but the classes, the split of train/test images, can easily be switched out for what you want.

0- Install the required libraries (OpenCV, Numpy, Jupyter Notebook, Pandas)

1- Download train-annotations-bbox.csv into this repository

2- Edit the Jupyter Notebook script's parameters to suit your needs

3- Run the script

Assuming you use the default parameters, you'll end with your directories and files looking like this:

• Object_Detection_DataPreprocessing.ipynb
• train-annotations-bbox.csv
• train-images-boxable.csv
• class-descriptions-boxable2.csv
• BP
  • classCSVs
    • arm_img_url.csv
    • beard_img_url.csv
    • ...
  • data
    • arm
      • (arm images)
    • beard
      • (beard images)
    • ...
  • test
    • (test images)
  • train
    • (train images)
  • annotations.txt
  • test.csv
  • train.csv

annotations.txt is what you will want to use for training your model(s) on the custom dataset. These annotations are in the format (img_name, x1, y1, x2, y2, label) which are more acessible for training.

Credit to RockyXu66's work on his Faster RCNN repo for Open Images in which this script is based on.