Under the hood: Building accessibility tools for the visually impaired on Facebook

This post details the development of automatic alternative (alt) text for photos on Facebook for iOS, utilizing object recognition technology powered by a deep convolutional neural network. It describes the process of selecting approximately 100 prominent and accurately detectable concepts, constructing natural-sounding sentences from detected objects and scenes, and the challenges of balancing information richness with accuracy and social context. The goal is to provide a scalable, automated solution to improve accessibility for visually impaired users.