【师者】Xavier Lagorce: 寻找自我的航程

The scientific journey is a long and tumultuous one but is as equally rewarding. In your late studies and your first steps in research you will probably feel a little lost, but it is a sign that you are reaching into the unknowns of the world. Paving the way, stone after stone, for the next generation of scientists to follow you into, yet unexplored, scientific areas.

This is the kind of research that always attracted me. Challenging old paradigms and exploring new topics that only a few had seen before. When looking at the history of research or famous researchers talking about their work, breakthroughs often look simple and easy but they are always the outcome of hard, and sometimes life-long, work.There is no single way of looking at research and any path can always bend towards a new destination. I strongly encourage you to explore, try and find your way through research. Only the joy of working on something you love and believe in can bring success in our fields where competition is everywhere. Find this, and research won't be work.

Dr Lagorce's research revolves around Neuromorphic Engineering, especially event-based cameras and Neuromorphic computing.

Neuromorphic Engineering is a new application field that finds its roots a few decades ago in the US. The basic idea is to notice that the underlying processes that rule the behaviors of neurons in our brain are quite similar to the equations of our modern transistors. Instead of building typical digital circuits, one should then be able to build electronic circuits quite similar to the brain but with much faster dynamics, if only we can understand how neurons interconnect.

Event-based cameras are a first approach to this which aims at reproducing one of the most accessible part of the brain: our eyes and especially the retina. They take inspiration in those biological devices to change how we acquire the visual information into a digital system. All pixels become independent, like the photoreceptors of our eyes, looking for changes in the light they receive and notifying the brain when that happens. Frames no longer exist, only events. Time is becoming the most valuable information.

Those sensors allow to acquire very sparse data because they only sample updates to the visual scene. They can sense very fast changes as well as adapt to very diverse light conditions. Those properties make them ideal to be integrated into robots or low-power systems.

More and more companies are building such kind of sensors and they are becoming a mature available consumer product. But the shift from frames to events also calls for a paradigm shift in how we process this data in computer vision. Those new approaches, and the applications they unlock are at the center of Dr Lagorce's research.

Those algorithms are also a direct foray into what computing with events or spikes can be, thus paving a way to explore new computing architectures that are no longer bounded by the large memory storages and transfers required by frames. Such non Von Neumann architectures, like neuromorphic computing systems and chips are one of the possible future of computing as we know it today.

Event-based Vision for Robotics

The most direct applications of the team are geared towards robotics. Event-based sensors offer a very sparse and precise way of measuring dynamic information such as motion. Comparing this data to the images that we were used to is a total game changer. Instead of a few frames per seconds where objects are jumping around, you now get a fully dense sampling of moving objects or edges through every pixel they cross. Algorithms can now process when things happen instead of where they happen.

This offers broad new perspectives into the perception of speed and motion and several of our research projects are looking into optical flow and navigation algorithms such as visual odometry or Simultaneous Localization And Mapping (SLAM).

We are also interested in new ways to represent those motions, going for more continuous representations to replace typical time sampled trajectories in order to better describe the evolution of objects' position through space as well as their characteristics such as speed and acceleration.

Active Light Systems

Since Event-based cameras are very good at seeing fast and fleeting things, we are also interested in active systems where we now control the light sources. This encompasses a broad range of applications where the light is reflected by surfaces, like structured light for 3D reconstruction or where the light sources are directly observed by the cameras like Visible Light Communication (VLC).

VLC is a quite interesting topic that can move object markers beyond the typical limits of QR-codes or other similar 2D patterns used today in computer vision.

Spatial Monitoring

Another key limitation of today's computer vision systems, especially related to artificial intelligence, is the amount of computing resources (and power) they require to operate. This limits the scaling possibilities of such systems because of the costs they would create, both in term of hardware but also in energy consumption.

Scaling such AI systems to simpler hardware and low-power platforms is a key challenge for deploying intelligent systems at large scales like cities and is a very rewarding new research topic for event-based cameras as smaller and cheaper sensors become available.

I am the father of 2 small kids that are delightfully occupying a lot of my spare time out of research and helping me bring patience as a core value to my teaching. When I have some free time left I enjoy traveling, listening to music, watching a good movie/TV show, killing time on a golf course or trying to brew some nice coffee (some says I'm kind of a coffee geek).

I also quite enjoy photography and scuba diving (which is considered an extreme sport in France with its long pioneering history in the field)... or sometimes doing both at the same time.