In a recent study, MIT researchers investigated the concept of peripheral vision in AI models. Unlike humans, AI lacks the ability to see objects in its visual periphery. By training computer vision models with an image dataset that simulates peripheral vision, the researchers found that the models improved in detecting objects but still fell short of human performance.
Simulating peripheral vision
Humans use their sharp, central vision called the fovea to focus on details, while the rest of their field of vision is considered the visual periphery. To mimic this process in AI, the researchers developed a technique based on the human texture tiling model, which simulates visual information loss in the periphery. By transforming images to represent this loss, they created a dataset to train computer vision models.
Peculiar performance
When humans and models were tested on an object detection task using transformed images, humans outperformed the machines. Even though training models with the dataset improved their performance, they still struggled to detect objects in the far periphery. Further research is needed to understand these differences and develop AI systems that can predict human behavior in the visual periphery.
This study sheds light on the importance of understanding peripheral vision in AI and its potential applications in driver safety and human-computer interaction. By exploring these differences, researchers aim to bridge the gap between AI models and human vision capabilities.