Meta has released a new tool designed to detect racial and gender bias in computer vision systems.
Many computer vision models show systemic bias against women and people of color. Meta hopes that improved tools will allow developers to better detect and fix bugs.
The company provides researchers with access to FACET (Fairness Evaluation in Computer Vision), a tool for evaluating the performance of computer vision models based on various characteristics, including skin color.
FACET is based on more than 30,000 images of 50,000 people, which experts divide into different demographic categories.
The company notes that the tool can be used to answer questions, such as whether the engine is more effective in identifying skaters when the skater's gender is male, or whether the system is more effective than before in identifying people with light skin compared to people with dark skin. . These problems are exacerbated when a person's hair is curly rather than straight.
Meta has relicensed the DINOv2 computer vision model under the Apache 2 open source license, allowing commercial use.
A series of DINOv2-based dense prediction models for semantic image segmentation and monocular depth estimation have also been released, providing developers and researchers with great flexibility to explore the model's capabilities in downstream tasks.
Meta says the move addresses the difficulty of measuring fairness in image processing, which often fails due to potential misnomers and demographic biases.
“Measuring fairness standards in computer vision is extremely difficult because the risk of misclassification is real and people using these AI systems can have good or bad experiences,” Meta said in a blog post. System complexity is not important. Same tasks, but combined. “Population.”
The company said it continues to develop AI systems while recognizing the potential harmful impacts of these technological developments on historically marginalized and underrepresented communities while seeking to address those impacts.