This is a reduced version of the original 500gb dataset. In the small version i sampled 5000 train images and 500 train/val images. This should be a handy drop-in replacement to create a dev environment and test stuff out.
Download Small OpenImagesV4 (2GB): https://drive.google.com/file/d/14nukC03LSp37Qw8XOAdNOsJNJ_Px2fb2/view?usp=sharing
The naming is unchanged as here: https://storage.googleapis.com/openimages/web/download.html
~tree --filelimit=10 small_openimages/
small_openimages/
├── bbox
│ ├── test-annotations-bbox.csv
│ ├── train-annotations-bbox.csv
│ └── validation-annotations-bbox.csv
├── challenge2018
├── class-descriptions-boxable.csv
├── imageIds
│ ├── test-images-with-rotation.csv
│ ├── train-images-boxable-with-rotation.csv
│ └── validation-images-with-rotation.csv
├── image-labels
│ ├── test-annotations-human-imagelabels-boxable.csv
│ ├── train-annotations-human-imagelabels-boxable.csv
│ └── validation-annotations-human-imagelabels-boxable.csv
├── test [499 entries exceeds filelimit, not opening dir]
├── train [4951 entries exceeds filelimit, not opening dir]
└── validation [490 entries exceeds filelimit, not opening dir]
Be aware that is is a random sample so it might happen that it does not reflect the original dataset distribition. And it might not contain all the labels!
Use the notebook to customize what you need. Different sampling? just downloading the label files? checkout the notebook
To download the whole dataset(all images) follow https://github.com/cvdfoundation/open-images-dataset#download-images-with-bounding-boxes-annotations
You can use the notebook to download the label files and set up the dataset structure