Our mission is to develop the body of theoretical knowledge and algorithms necessary to robots – and intelligent systems in general – to learn autonomously about objects in an open-ended manner. That implies using tools from the world of machine learning, computer vision multimodal signal processing and analysis, data visualization and mining. Although intelligent embodied systems are the main application driving research in VANDAL, other applications are non-invasive control of prosthetic hands, scene understanding and automatic geolocalization.
OUR MAIN AREAS OF RESEARCH
Let's think about three different images of a sheep: the first one is a photo taken during a walk; the second is a painting and the last a child's drawing. Human beings have no difficulty in understanding that a sheep is depicted in all of them. But when it comes to neural networks, it is not that easy. This is where domain adaptation comes into play: its goal is to get models capable of solving tasks in a target domain different from the one they were trained on, i.e. the source domains.
Humans have the capability to learn effortlessly new concepts over time. Emulating this capability in neural networks has demonstrated to be hard since they tend to forget what they learned in the past while learning novel concepts. The goal of incremental learning is studying how neural networks can be extended over time, making them lifelong learners.
First Person Action Recognition
First person action recognition is an increasingly researched topic because of the growing popularity of wearable cameras and for its central role in real-world egocentric vision applications, from wearable sport cameras to human-robot interaction or remote assistance. The wearable camera is usually mounted on the person’s head and the goal is to recognize egocentric actions resulting from the interactions with objects, such as “cutting a tomato” or “preparing a meal”.