Tools for downloading images and corresponding annotations from Google's OpenImages dataset.
The openimages package contains a download module which provides an API with
two download functions and a corresponding CLI (command line interface) including
script entry points that can be used to perform downloading of images and
corresponding annotations from the OpenImages dataset.
openimages.download.download_images for downloading images only
For example, to download all images for the two classes "Hammer" and "Scissors" into the directories "/dest/dir/Hammer/images" and "/dest/dir/Scissors/images":
from openimages.download import download_images
download_images("/dest/dir", ["Hammer", "Scissors",])openimages.download.download_dataset for downloading images and corresponding
annotations
For example, to download all images and corresponding annotations in PASCAL
VOC format for the two classes "Hammer" and "Scissors" into the directories
"/dest/dir/Hammer/[images|pascal]" and "/dest/dir/Scissors/[images|pascal]":
from openimages.download import download_dataset
download_dataset("/dest/dir", ["Hammer", "Scissors",], annotation_format="pascal")Two Python script entry points are installed when the package is installed into
a Python environment, corresponding to the public API functions described above:
oi_download_dataset and oi_download_images. These commands use the follwing
options:
| Option | Required | Description |
|---|---|---|
| --base_dir \<dir> | yes | directory into which images and annotations will be downloaded, with each class label having a separate subdirectory containing an "images" subdirectory for image files and (for annotated datasets) an \<annotation_format> subdirectory for annotation files |
| --labels \<label1> [\<label_2> ...] | yes | space-separated list of class labels, at least one required, multi-word labels with spaces must be quoted |
| --format \<annotation_format> | for annotated dataset yes, not applicable for images only | required for downloading an annotated dataset, currently supported format specifiers are "darknet" and "pascal" |
| --meta_dir \<dir> | no, but usually recommended | directory into which the metadata files specifying annotations and class labels are downloaded (if not already present) or read from (if present) |
| --exclusions \<file> | no | text file containing image file IDs, one per line, for images to be excluded from the final dataset, useful in cases when images have been identified as problematic |
| --limit \<int> | no | the upper limit on the number of images to be downloaded per label class |
| --include_segmentation | no | if included then segmentation data will be included in the resulting dataset, only valid in conjunction with "pascal" output annotation format |
If you'll use these commands more than once then it's imperative to utilize the
--meta_dir option that specifies where to save the (rather large) metadata files containing
bounding box information etc., as this will save you from having to redownload this
large file in subsequent usages.
Download images and PASCAL format annotations for the class labels "Scissors" and
"Hammer", limiting the number of images to 200 and storing the CSV files under
~/openimages (reading the CSV files from there if they already exist):
$ oi_download_dataset --meta_dir ~/openimages --base_dir ~/openimages --labels Scissors Hammer --format pascal --limit 100Download images only for the class label "Scissors", limiting the number of images
to 100 and storing the CSV files under ~/openimages (reading the CSV files from
there if they already exist):
$ oi_download_images --meta_dir ~/openimages --base_dir ~/openimages --labels Scissors --limit 100