The 2024 Conference on Computer Vision and Pattern Recognition, sponsored by the IEEE Computer Society and the Computer Vision Foundation, is the leading annual conference for computer vision sciences and the premier destination for cutting-edge research in related computing fields such as artificial intelligence, deep learning, and augmented/virtual reality.
This year’s conference included the inaugural Workshop on AI for Streaming, which focuses on “unifying new streaming technologies, computer graphics, and computer vision, from the modern deep learning point of view.” WNCG’s Dr. Al Bovik presented the keynote speech, titled “On the Visual Quality of Pictures, Games, and GenAI."
Decades of research in computer vision have led to the current state of streaming we enjoy today. Highly photorealistic digital images and videos can be streamed with greater efficiency, and as image and video viewing accounts for up to 80% of internet traffic, this reduced bandwidth consumption translates to a lower carbon footprint for internet usage. Major global players in the streaming industry have taken advantage of this research to improve the overall quality and user experience of online gaming and streaming.
In an interview with WNCG staff, Bovik described some new challenges facing this technology due to the emergence of AI generated images. Existing algorithms were designed around visual perception of real objects, but computer-generated imagery does not exist in nature.
“Visual distortion of real images is quantifiable based on models of neural responses in the visual brain. With GenAI, there is no reference point or usual concept of distortion.”
Not only will engineers have to accommodate for these new parameters, but a new GenAI-centric model would have to account for the intention of the creator: how to make AI-generated images and videos realistic and convincing, despite some images and graphics being intentionally other-worldly.
Bovik’s keynote discussed whether existing algorithms can be adapted to meet these new requirements, including a new approach to quality prediction of online gaming called GAMIVAL, among other new approaches to measure the quality and “believability” of GenAI images.