What is FHIBE?
The Fair Human-Centric Image Benchmark* (pronounced Fee-Bee) is the first publicly available, globally diverse, consent-driven dataset built specifically to evaluate fairness in human-centric computer vision tasks. While AI vision powers everything from autonomous vehicles to smartphones, most datasets still suffer from bias, lack of diversity, and ethical concerns around consent and compensation.
Why a Fairness Benchmark?
Bias evaluation is essential for ethical AI. Yet most evaluation datasets fall short: lacking consent, or detailed annotations. FHIBE fills that gap as a responsibly curated, consent-based dataset designed to advance fairness in computer vision. How FHIBE is different:
Consent
Every dataset participant gave informed, revocable consent. Data collection is designed to comply with data protection laws.
Compensation
All image subjects, annotators, and QA reviewers were fairly paid at or above local minimum wage. Participants can withdraw their data at any time without affecting compensation.
Diversity
FHIBE maximizes diversity across demographics, appearance, poses, and environments. It includes detailed data on age, pronouns, ancestry, skin tone, eye color, hair type, and visible markers to support granular bias analysis.
Comprehensive Annotations
FHIBE includes Pixel-level labels, 33 keypoints, and 28 segmentation categories—linked to anonymized annotator IDs.
Real-World Utility
FHIBE is designed for fairness evaluation across pose estimation, face/body detection, segmentation, synthesis, and vision-language models.
*FHIBE is strictly for fairness/bias purposes – FHIBE cannot be used for training, with the narrow exception of training bias mitigation tools.
A Fair Reflection - a short film
Sony’s Quest to Make AI Models Work for Everyone
Many ethical issues with computer vision start at the data layer when data is often not collected in an ethical and responsible way. Ethical data collection is an extremely challenging issue and FHIBE is the first dataset of its kind, a new benchmark for the community. It’s important that the whole industry builds AI on a foundation of ethical data. Watch our short film to learn more.
Dataset Details
More than 10,000 Images from Nearly 2,000 People
FHIBE includes 10,318 diverse, consented images—each featuring one or two people, captured across 81 jurisdictions using varied environments and camera types. Most participants contributed around six images each, all provided directly by the individuals depicted.
FHIBE includes two-face focused assets:
- Unaligned face-cropped dataset
- Aligned face-cropped dataset
To protect privacy, all non-consensual individuals have been fully anonymized, and all personally identifiable information (PII) has been redacted, minimizing re-identification risk.
Comprehensive Annotations and Rich Metadata
- Self-reported image subject attributes (not guessed)
- Face and person bounding boxes
- 33 body and facial keypoints
- 28 segmentation categories
- Anonymized annotator IDs
Evaluating Fairness in Computer Vision with FHIBE
FHIBE supports fairness evaluation across human pose estimation, face and body detection, segmentation, verification, image editing and synthesis, and vision-language models.
Human Pose Estimation
FHIBE supports the evaluation of human pose estimation models by providing 2D keypoint annotations (x, y coordinates) that capture the geometric structure of human bodies and faces. The dataset includes 33 keypoints. Keypoints are localized for major landmarks, such as the right eye inner corner, nose, right hip, and left foot index.
Face and Body Detection
FHIBE enables the evaluation of face detection and person detection models by providing bounding boxes with precise (x, y) coordinates for each image subject. These annotations allow for accurate localization of both the human body and face within an image.
Segmentation
FHIBE enables the evaluation of segmentation models by providing pixel-wise annotations that partition the human body and face into 28 distinct segmentation categories (e.g., face skin, inner mouth, bag, right shoe). Each pixel is assigned a label, ensuring precise differentiation of anatomical regions and accessories.
Identity Verification
FHIBE supports the evaluation of face verification models, which determine whether two images show the same person. To ensure reliable comparisons, FHIBE includes an aligned face-cropped dataset, providing standardized images for testing identity verification systems.
Image Editing and Synthesis
FHIBE supports the evaluation of image restoration and super-resolution models, assessing their ability to enhance or reconstruct human-centric images. Image restoration recovers degraded images by removing noise, correcting distortions, and restoring lost details. Super-resolution sharpens low-resolution images, improving clarity while preserving the subject’s identity and attributes.
Vision-Language Models
FHIBE enables the evaluation of vision-language models (VLMs) in image comprehension and recognition tasks, including visual question answering (VQA). By providing diverse, annotated human-centric images, FHIBE helps assess bias in how VLMs reason about people.
Frequently Asked Questions
FHIBE is a Project of Sony AI
Over the course of three years, a global team of Sony AI researchers, engineers, policy specialists, and project managers worked to develop the rigorous procedures for data collection, annotation, and validation that became FHIBE. Their work was further supported by legal, privacy, IT, and QA specialists. The Sony AI Ethics research team is leading the way towards more ethical AI that protects the interests of AI users and creators by pushing the industry to ensure AI technologies are fair, transparent, and accountable.
Sony AI, a division of Sony Research, was founded as a subsidiary of Sony Group on April 1, 2020, with the mission to "Unleash human imagination and creativity with AI."
Learn more about Ethics Research at Sony AI.