A team from MIT has created an image dataset aimed at simulating peripheral vision in machine learning models, a characteristic which AI typically lacks. This could improve the models’ ability to recognise approaching threats and predict whether a human driver would spot an oncoming object. In experiments, these models improved in terms of hazard detection, though not as significantly as humans. The size of objects and the amount of visual clutter appeared to have little effect on AI’s ability to notice them. This research could help develop models which view the world similarly to humans. Lead Author, Anne Harrington MEng ’23, said understanding the features in a visual scene which make our eyes move to collect more information could help simulate peripheral vision. The research showed that peripheral vision can play a significant role in human interaction with machines such as cars, robots and user interfaces.
The texture tiling model was used to simulate peripheral vision in humans, with the team from MIT adapting the model so it could transform images in a way that doesn’t necessitate knowing in advance where the person or AI would focus their eyes. The modified technique was used to create a large dataset of transformed images which appear more textured in some places, to demonstrate the loss of detail that happens when a human looks further into the periphery.
The models and study subjects were shown pairs of transformed images which were identical except for a target object located on the periphery of one and they were asked to identify the image with the target object. The study found that the models and humans performed differently in this task, suggesting the models may not use context in the same way as humans for these detection tasks.
Despite the strides made in AI, the performance of the neural network models did not match human performance in detecting objects in the periphery. This shortcoming could prompt additional AI research into machine learning models which closely approximate human vision. The dataset, which is publicly available, is meant to encourage more related research. Supported by the Toyota Research Institute and the MIT CSAIL METEOR Fellowship, this work could result in AI systems that alert drivers to unseen hazards.
