Contributors

Trained models

• FasterRCNN Inception ResNet V2 and SSD Mobilenet V2 object detection model (trained on V4 data). Model checkpoints

• FasterRCNN Inception ResNet V2 object detection model (trained on V2 data). Model checkpoint, evaluation protocol, and inference and evaluation tools are available as part of the Tensorflow Object Detection API.

• Resnet 101 image classification model (trained on V2 data): Model checkpoint, Checkpoint readme, Inference code.

The models are released under an Apache 2 license.

Community Contributions

• Jason Kuen and co-authors shared a pretrained object detector for 5K categories, trained using both Open Images box-levels and image-level annotations. Training code is also available. From their ICCV 2019 paper.

• The Conceptual Captions Challenge uses 1,000 Open Images images for their open test set to evaluate conceptual captioning algorithms.

• A team from the Georgia Institute of Technology and Facebook AI Research released nocaps, which augments the Open Images val and test sets with 166,100 natural language captions describing 15,100 images.

• Vittorio Mazzia and Angelo Tartaglia wrote a ToolKit to help you download subsets of images from Open Images V4 filtering by class, attributes, etc.

• The team at Algorithmia created an in-depth object detection tutorial that walks through how to use the provided bounding box annotations to create a useful object detection model with Tensorflow.

• Dan Nuffer offers helper code to retrieve the images at Open Images dataset downloader. It is a program built for downloading, verifying and resizing the images and metadata. It is designed to run as fast as possible by taking advantage of the available hardware and bandwidth by using asynchronous I/O and parallelism.