Open Images

Open Images
Name	Open Images
Subject	dataset
Creator	Google Research
Released	2016
Format	images with bounding boxes, segmentation masks, labels, visual relationships
License	varied (CC BY, CC0, custom)

Open Images

Open Images is a large-scale, publicly distributed image dataset assembled by researchers at Google Research to support computer vision research and benchmarking. It aggregates millions of images with dense annotations designed for tasks such as object detection, image classification, instance segmentation, and visual relationship detection. The dataset has been used alongside datasets like ImageNet, COCO (dataset), PASCAL VOC, and Places (dataset) in work from institutions including Stanford University, Massachusetts Institute of Technology, Carnegie Mellon University, and University of Oxford.

Overview

Open Images provides a broad corpus of imagery drawn from web sources and curated for research by teams at Google Research and collaborators at organizations such as Cornell University, ETH Zurich, University of Toronto, and Facebook AI Research. The collection emphasizes scale and annotation richness to enable training of deep models developed in communities around frameworks like TensorFlow, PyTorch, and libraries maintained by groups including OpenAI, DeepMind, and Microsoft Research. It complements historical benchmarks produced by projects associated with ImageNet creators at Princeton University and evaluation tracks run by forums such as CVPR, ICCV, and ECCV.

Dataset Composition and Annotations

The dataset includes millions of images annotated with multi-label image-level labels, thousands of bounding-boxed object instances, and segmentation masks. Annotators and research partners from institutions like Appen, teams from Google Brain, and crowd-sourcing efforts tied to platforms associated with Amazon Mechanical Turk contributed annotations. Annotation taxonomies reference ontologies and vocabularies developed at Wikidata and linked to identity systems used by projects from Mozilla and Creative Commons. Benchmarks using Open Images often compare performance against models trained on corpora curated by labs such as DeepMind, Facebook AI Research, NVIDIA Research, and university groups at Caltech, University of California, Berkeley, and University of Washington.

Versions and Releases

Open Images has undergone multiple releases that expanded labels and refined annotations, with major versions announced by research teams at Google Research and presented at conferences like NeurIPS, ICCV, and CVPR. Successive releases increased the number of categories and added tasks such as visual relationship detection, instance segmentation, and panoptic annotations, following trends set by predecessors like COCO (dataset) and initiatives at Kaggle. Academic groups at University College London, Imperial College London, EPFL, Max Planck Institute for Informatics, and Tsinghua University have cited specific Open Images releases in papers on model generalization and transfer learning.

Licensing and Usage Terms

Licensing for images in the dataset is heterogeneous and was curated to permit research reuse under terms linked to providers such as Creative Commons licensors. Usage policies reflect considerations familiar to repositories administered by Wikimedia Foundation, policies debated in venues like ICLR and legal discussions referencing frameworks from Creative Commons and national bodies such as European Commission research initiatives. Commercial and non-commercial reuse is governed by metadata and license fields included with images; users from labs like Google Research, Microsoft Research, IBM Research, and startups in the Mozilla ecosystem routinely inspect those metadata prior to model training or dataset redistribution.

Applications and Impact

Open Images has been applied to object detection, segmentation, relationship modeling, zero-shot learning, and dataset bias studies by researchers at Stanford University, Massachusetts Institute of Technology, Carnegie Mellon University, University of Oxford, University of Toronto, MIT-IBM Watson AI Lab, and industry groups at Google, Microsoft, Amazon, Facebook, and Apple. It has influenced competitions and leaderboards hosted by Kaggle, challenge tracks at CVPR and ICLR, and industry benchmarks created by NVIDIA and Intel. Papers leveraging Open Images have intersected with research on architectures and methods developed by teams behind ResNet, Transformer (machine learning model), EfficientNet, and approaches from OpenAI and DeepMind.

Tools and Access Methods

Researchers access Open Images through programmatic download tools, cloud-hosted storage used by Google Cloud Platform and mirrors on platforms like Kaggle and institutional repositories at University of California, Berkeley and Stanford University. Tooling for annotation, visualization, and conversion interoperates with ecosystems maintained by TensorFlow, PyTorch, Detectron2, and utilities from projects like LabelImg, COCO API, and community toolkits authored by groups at Facebook AI Research and AWS (Amazon Web Services). Tutorials and reproducible code leveraging Open Images are distributed by academic labs including UC Berkeley AI Research, MIT CSAIL, University of Washington, and corporate research teams at Google and Microsoft Research.

Category:Image datasets