OpenImages

OpenImages
Name	OpenImages
Type	Image dataset
Creator	Google Research
Released	2016
Size	"≈9 million images (varies by release)"
License	"Creative Commons Attribution"
Website	"See Licensing and Access"

OpenImages OpenImages is a large-scale annotated image dataset curated for computer vision research and benchmarks. It was created by Google Research to support tasks such as object detection, visual relationship detection, and image classification, and has been used alongside datasets like ImageNet, COCO, PASCAL VOC, Cityscapes, and ADE20K in academic and industrial evaluations. The dataset has been cited in work associated with institutions such as Stanford University, Massachusetts Institute of Technology, University of Oxford, Carnegie Mellon University, and companies including Facebook AI Research, Microsoft Research, and DeepMind.

Overview

OpenImages originated as a response to the need for broader, more diverse visual data than offered by prior collections such as ImageNet and PASCAL VOC. It emphasizes dense annotations and multi-label images to support complex tasks referenced by research groups at Google Research and conferences like NeurIPS, CVPR, ICCV, and ECCV. The dataset has versions and challenge tracks used in venues including the Google AI Challenge and datasets compared in papers from labs at Berkeley AI Research, ETH Zurich, University of Toronto, and Tencent AI Lab.

Dataset Composition

The corpus contains millions of images sourced under permissive terms similar to collections like Flickr Commons and previously released corpora connected to projects at Getty Images licensing programs; its composition spans categories that echo taxonomies used in ImageNet and label sets comparable to COCO and OpenCV examples. Releases include hierarchical class sets that intersect with ontologies derived from resources such as WordNet and label schemas evaluated by teams at Yahoo Research and Adobe Research. The dataset provides bounding boxes, instance masks in subsets, and image-level labels with class lists overlapping concepts familiar to research groups from Oxford Visual Geometry Group and datasets employed by Google Cloud AI customers.

Annotation and Labeling

Annotations were produced using workflows influenced by practices at Amazon Mechanical Turk studies, professional annotators used in projects from Labelbox partners, and quality-control procedures similar to those described by researchers at Microsoft Research AI. Label taxonomies reference canonical nodes in WordNet and align with class definitions used in challenges at Kaggle and evaluation suites from WILDLABS. Annotations include bounding-box coordinates for objects, visual relationship triplets that echo relationships analyzed by teams at Facebook AI Research (examples: person-holding, dog-on), and attribute labels used in experiments by Toyota Research Institute and NVIDIA Research.

Usage and Applications

Researchers and engineers employ the dataset for tasks cited in publications from Stanford Vision Lab, MIT CSAIL, Carnegie Mellon University Robotics Institute, and Imperial College London to train and evaluate models such as variants of Faster R-CNN, Mask R-CNN, YOLO, SSD (Single Shot MultiBox Detector), and transformer-based architectures like ViT and Swin Transformer. OpenImages has been used in applied projects at companies including Google, Amazon, Apple, Microsoft, and Tesla for object detection, scene understanding, and multimodal learning that integrates methods from groups at OpenAI and DeepMind. It supports cross-dataset transfer-learning studies comparing performance with benchmarks from COCO, ImageNet, Cityscapes, and domain adaptation research in papers presented at ICLR and NeurIPS.

Licensing and Access

Distribution follows permissive Creative Commons attribution-style terms similar to licensing patterns visible in datasets released by Wikimedia Foundation and media collections referenced by Creative Commons. Access and download portals have been integrated with cloud services like Google Cloud Storage and academic mirrors used by repositories linked to Zenodo and Kaggle Datasets. Usage in commercial products and academic projects has been discussed in guidance from legal teams at Stanford Law School clinics and policy groups at Electronic Frontier Foundation when aligning dataset use with compliance considerations such as those addressed by GDPR and institutional review boards at universities like Harvard and Yale.

Evaluation and Benchmarks

The dataset provides standard splits and evaluation protocols that have been adopted in benchmark suites compared across leaderboards maintained by communities at Papers with Code, challenge tracks at CVPR Workshops, and competitions hosted on Kaggle. Metrics commonly reported by teams at Microsoft Research and academic groups include mean average precision (mAP), intersection over union (IoU), and per-class recall used in comparisons with models developed at Facebook AI Research, Google Brain, OpenAI, and research labs at NVIDIA. Leaderboard results using OpenImages variants are regularly cited in publications from University of Cambridge, Princeton University, Cornell University, and industrial research groups in evaluation papers at ICCV and ECCV.

Category:Computer vision datasets