Why Egocentric Vision matters to me

There is a fundamental difference between watching someone do something and seeing the world through their eyes.

In traditional computer vision, the camera typically observes a scene from the outside. A person cooks, assembles something, or opens a drawer.

In egocentric video, the camera moves with the person.

You see their hands, the objects they interact with, where they are looking, and an environment that constantly shifts with their movements.

That is what interests me most.

Egocentric vision is not only about recognising a person’s actions. It can play a central role in systems that must understand what is happening around them and anticipate what is likely to happen next.

Consider a robot that collaborates with a person and needs to anticipate their next move. Or consider a production line where a system monitors a process, determines which stage a worker has reached, and detects deviations or errors. The same applies in an operating room, where a first-person view can help identify the phases of a procedure, the instruments being used, and the context in which each action takes place.

The question is no longer simply:

"What is the person doing?"

but:

"What has happened so far, and what are they likely to do next?"

This is known as action anticipation, and it is the main focus of my research.

To predict a future action, a model must draw on information from the past: movements, objects, context, interactions, and potentially multiple modalities.

This raises questions that I find far more interesting than a simple classification problem.

How much of the past should the model remember? What information is genuinely worth retaining? How can it combine visual appearance, motion, audio, and object information when each evolves on a different timescale? And how can we ensure that an anticipation model does not use information from the future, even accidentally?

This year, we also worked on EgoShift, examining what happens when a model is deployed across different users and environments and must adapt to a new domain. At the same time, as part of my PhD, I am building the research infrastructure needed to run such experiments with different visual and multimodal representations, proper temporal alignment, and reproducible evaluation.

The more I work on egocentric vision, the more I believe its real value lies in a broader question:

how an intelligent system observes the world over time, remembers what matters, and uses the past to understand what is likely coming next.