Datasets

SCIN Crowdsourced Dermatology Dataset

The SCIN dataset contains 10,000 images of dermatology conditions, crowdsourced with informed consent from US internet users. Contributions include self-reported demographic and symptom information and dermatologist labels, as well as estimated Fitzpatrick skin type and Monk Skin Tone.

BC-Z Demonstration Dataset

Episodes of a robotic arm performing 100 different manipulation tasks. Data for each episode includes the RGB video, the robot's end-effector positions, and the natural language embedding. Episodes were gathered using teleoperation via a VR controller.

Crossmodal-3600

Crossmodal-3600 is a geographically diverse dataset of 3600 images each of them annotated with human-generated reference captions in 36 languages.

EditBench

EditBench is a comprehensive diagnostic and evaluation dataset for text-guided image editing.

Auto-Arborist

The Auto Arborist dataset is a multiview fine-grained visual categorization dataset that contains over 2 million trees belonging to over 300 genus-level categories in 23 cities across the US and Canada built to foster the development of robust methods for large-scale urban forest monitoring.

UGIF

A corpus of Android how-to queries (speech and text) in eight languages along with how-to instructions in English paired with sequences of UI screens and actions as the how-to is completed by human annotators on Android devices with different UI language settings.

KIP Distilled Datasets

These are distilled datasets derived from MNIST, Fashion-MNIST, CIFAR-10, CIFAR-100, and SVHN using infinitely wide convolutional networks. Sample result: Over 64% test accuracy on CIFAR-10 achieved using only 10 images.

Lens Flare

High-quality RGB images of typical lens flare against a black background. Among them, ~2k are captured with a typical smartphone camera, and ~3k are simulated computationally.

WIT Wikipedia-based Image Text Dataset

WIT is a large Multimodal, Multilingual dataset created using Wikipedia data. WIT contains ~37M+ image-text example sets across 108 languages. This makes WIT one of the biggest image-text dataset publicly available in addition to it being very entity-rich and providing contextual information.

Google Open Images Mutual Gaze dataset

This dataset consists of images along with annotations that specify whether two faces in the photo are looking at each other. This dataset is intended to aid researchers working on topics related to social behavior, visual attention, etc.

Open Images

A dataset consisting of ~9 million URLs to images that have been annotated with labels spanning over 6000 categories.

DQN Replay dataset

An offline RL dataset on Atari 2600 games based on the logged replay data of a DQN agent comprising 50 million (observation, action, reward, next observation) tuples per game.

Google Landmarks Dataset v2

The Google Landmarks dataset (GLDv2) is a large-scale benchmark for fine-grained instance-level recognition. It contains over 5M images of natural or human-made landmarks and has protocols for evaluating object recognition and image retrieval.

HDR+ Burst Photography Dataset

An archive of full-resolution raw image bursts over a wide range of scenes, along with the results from Google's HDR+ camera software for comparison.

Open Images Extended - Crowdsourced

Additional imagery sets to the main Open Images dataset, to improve its diversity (geographic, cultural, demographic, subject matter, etc). Currently composed of ~478K images contributed by users of the Crowdsource app.