Teaching AI to Understand Videos from a First-Person View

When we wear a camera on our body, like on a pair of glasses, it can capture what we're doing from a first-person point of view. However, this egocentric view is limited because it only shows what's happening from one angle.

To get a better understanding of what's happening, we need to combine information from different viewpoints, like what the camera sees and what's happening in the environment. This is where multi-teacher distillation comes in - it's a way of teaching an AI model by combining the knowledge of multiple other models.

The problem is that these models might have different architectures or ways of understanding the world, which can make it hard for the AI to learn. To solve this, the researchers introduced proxy models that act as translators between the different models. These proxies help the AI learn from the different models in a way that's consistent and easy to understand.

The researchers also developed a way to select which proxies to use for each piece of data, so the AI only learns from the most reliable and confident sources. This helps the AI learn faster and more accurately.