Real-time visual description program for people with low vision wins Best Paper at UIST 2024

CSE authors were recognized for their development of WorldScribe, a program that generates live visual descriptions to assist people who are blind or have low vision.
A hand holds a smartphone toward a desk with a laptop computer on top. A TV hangs on the wall over the desk, and a bookshelf stands against the wall nearby. Text descriptions on the phone screen read "workbenches," "shelf," "cabinets," and "TV."
As a user scans their phone camera around a room, WorldScribe will create brief audio descriptions of the objects recorded by the camera. Illustration credit: Shen-Yun Lai, used with permission.

A world of color and texture could soon become more accessible to people who are blind or have low vision through new software that narrates what a camera records. 

The tool, called WorldScribe, was designed by Anhong Guo, assistant professor of computer science and engineering at the University of Michigan, along with CSE PhD student Ruei-Che Chang and undergrad Yuxuan Liu. Their project, titled “WorldScribe: Towards Context-Aware Live Visual Descriptions,” will be presented at the ACM Symposium on User Interface Software and Technology (UIST), where it is one of just four papers to be recognized with a Best Paper Award. 

WorldScribe uses generative AI (GenAI) language models to interpret the camera images and produce text and audio descriptions in real time to help users become aware of their surroundings more quickly. It can adjust the level of detail based on the user’s commands or the length of time that an object is in the camera frame, and the volume automatically adapts to noisy environments like crowded rooms, busy streets and loud music. 

“For us blind people, this could really revolutionize the ways in which we work with the world in everyday life,” said Sam Rau, who was born blind and participated in the WorldScribe trial study.

“I don’t have any concept of sight, but when I tried the tool, I got a picture of the real world, and I got excited by all the color and texture that I wouldn’t have any access to otherwise,” Rau said. “As a blind person, we’re sort of filling in the picture of what’s going on around us piece by piece, and it can take a lot of mental effort to create a bigger picture. But this tool can help us have the information right away, and in my opinion, helps us to just focus on being human rather than figuring out what’s going on. I don’t know if I can even impart in words what a huge miracle that truly is for us.”

During the trial study, Rau donned a headset equipped with a smartphone and walked around the research lab. The phone camera wirelessly transferred the images to a server, which almost instantly generated text and audio descriptions of objects in the camera frame: a laptop on a desk, a pile of papers, a TV and paintings mounted on the wall nearby.

The descriptions constantly changed to match whatever was in view of the camera, prioritizing objects that were closest to Rau. A brief glance at a desk produced a simple one-word description, but a longer inspection yielded information about the folders and papers arranged on top.

The tool can adjust the level of detail in its descriptions by switching between three different AI language models. The YOLO World model quickly generates very simple descriptions of objects that briefly appear in the camera frame. Detailed descriptions of objects that remain in frame for a longer period of time are handled by GPT-4, the model behind ChatGPT. Another model, Moondream, provides an intermediate level of detail.

A cartoon on the left shows a man entering an office, and a thought bubble shows that he is seeking a laptop computer. The office has two desks with desktop and laptop computers on them. Another cartoon on the right shows the man holding a guide cane and a smartphone. Five speech bubbles show the phone's audio descriptions made from the WorldScribe app. They describe the location of several laptops around the office, as well as differences in color and logo between certain laptops.
When the user is moving slowly around the room, WorldScribe will use GPT-4 to create colorful descriptions of objects. When asked to help look for a laptop, the tool will prioritize detailed descriptions of any laptops in the room. Illustration credit: Shen-Yun Lai, used with permission.

“Many of the existing assistive technologies that leverage AI focus on specific tasks or require some sort of turn-by-turn interaction. For example, you take a picture, then get some result,” said Anhong Guo, an assistant professor of computer science and engineering and a corresponding author of the study.

“Providing rich and detailed descriptions for a live experience is a grand challenge for accessibility tools,” Guo said. “We saw an opportunity to use the increasingly capable AI models to create automated and adaptive descriptions in real-time.”

Because it relies on GenAI, WorldScribe can also respond to user-provided tasks or queries, such as prioritizing descriptions of any objects that the user asked the tool to find. Some study participants noted that the tool had trouble detecting certain objects, such as an eyedropper bottle, however.

Rau says the tool is still a bit clunky for everyday use in its current state, but says he would use it everyday if it could be integrated into smart glasses or another wearable device.

The researchers have applied for patent protection with the assistance of U-M Innovation Partnerships and are seeking partners to help refine the technology and bring it to market.

The research was funded by U-M.

Guo is also an assistant professor of information within U-M’s School of Information.

Study: WorldScribe: Towards Context-Aware Live Visual Descriptions (DOI to go live next week: 10.1145/3654777.3676375)

The original version of this story was published by Michigan Engineering and written by Derek Smith.

Explore:
Anhong Guo; Honors and Awards; Research News; e-HAIL