Meeting François Brémond — In Simple Words

Before this meeting, I read your INRIA page and the full STARS team page.

I looked at your Google Scholar: recent papers on skeleton action recognition, multi-object tracking, and behaviour analysis.

I read about Toyota Smarthome, CoBTeK, and the 3IA Côte d'Azur chair.

Everything I saw matches with my work. That is why I really wanted to meet you.

STARS builds systems that understand video: detect people, track them, and recognize what they do.

You go from pixels to semantics — from what the camera sees to what people mean.

You use deep learning, multi-sensor fusion, and spatio-temporal reasoning.

And you apply it to real problems: metro, airports, homecare, older adults with dementia.

You do activity recognition and behaviour analysis. I built one for breathing in pediatric ICU.

You work on fall risk and older adult monitoring. I just published a paper on fall detection.

You use multi-sensor fusion. My fall paper does bimodal fusion: camera plus IMU.

You use deep learning for action. I use deep learning for late fusion.

My PhD used Kinect RGB-D cameras to measure breathing of sick children, with no contact.

Point clouds, mesh reconstruction, volume by octree, and scene flow — all classic STARS topics.

Validated in a real pediatric ICU, on real babies, against the ventilator.

This work also became a US patent.

Two sensors: a camera that sees the person, and an IMU worn on the body.

Two deep networks: one for pose from video, one for motion from IMU.

Then a late-fusion decision: fall or not fall.

Published in Sensors (MDPI), 2025, with a follow-up at IEEE ICECS 2025.

Your CoBTeK team works with the Nice University hospital on older adults with dementia.

My PhD was in the Sainte-Justine pediatric hospital in Montreal, on children in respiratory distress.

We both know how to work with clinicians, with real, messy, real-world data.

That is a rare skill — and I want to bring it into your team.

I built a solid base in Canada: teaching, research, hospital work, and a US patent.

Today, I want a bigger, more structured environment — with more resources and more impact.

At STARS, I want to learn from multidisciplinary teams and contribute at a much larger scale.

This is not a break with my past — it is a natural progression, one clear step forward.

INRIA is one of the top research institutes in Europe, and STARS has a clear, ambitious mission.

Sophia Antipolis is a great science-and-tech hub — industry and academia side by side.

The 3IA Côte d'Azur chair puts your team at the centre of French AI.

For me, this feels like the right place at the right time.

1. Strong 3D vision and RGB-D camera skills (Kinect v2, Azure DK).

2. Hands-on experience in multi-modal fusion with deep learning.

3. Real ability to work with clinicians and turn ideas into a US patent.

4. Solid teaching experience — a real plus for supervising PhD students.

Newer transformer and self-attention methods for video understanding.

Long-term activity mining and self-supervised learning.

How to run large European projects from proposal to results.

Every day working with a team of world-class researchers.

My first hope: a postdoc or a researcher position in the STARS team.

Topics I would love to work on with you: multi-modal action recognition, fall risk and behaviour of older adults, deep learning with 3D sensors.

I am ready to move — and I bring experience, energy, and clear research questions.

Thank you for taking the time to meet me.

I did not come with one idea — I came with four. But the one I would lead is called BehaviourWatch-3D.

The goal: watch older adults at home, without contact, and detect the early signs of decline — weeks before the clinician sees them.

The approach: one multi-modal foundation model trained on RGB, depth, and wearable signals together.

The output: weekly clinical reports, written by a vision-language model, sent to the doctor.

M1. Foundation model pre-training: self-supervised on RGB + depth + IMU, with masked auto-encoding and cross-modal contrastive learning — the 2026 state of the art.

M2. Contactless physiology: breathing (from my PhD), heart rate via rPPG, gait, posture — all from one depth camera.

M3. Behaviour understanding: activities of daily living (Toyota Smarthome style), agitation and apathy (CoBTeK).

M4. Personalised decline detection: a per-person self-supervised anomaly model — alert when the baseline shifts.

Year 1 (2026): pre-train the foundation model on public + Toyota Smarthome + CoBTeK data. First paper (CVPR / NeurIPS class).

Year 2 (2027): decline prediction, clinical validation at CoBTeK, vision-language reports. Second paper + patent extension.

Funding fit: 3IA Côte d'Azur, Horizon Europe (EU4Health, EIC), ANR PRC.

Real impact: earlier intervention in dementia care, fewer hospitalisations, better quality of life for families.

Idea 2 — Fall-Ahead-3D. Not just detecting falls — predicting them, hours to seconds ahead, from subtle gait, posture and breathing shifts. Extends my IEEE ICM 2026 paper on camera + IMU fusion.

Idea 3 — MedActionLLM. A vision-language model grounded on medical guidelines (RAG on SNOMED / ICD-10) that writes short, verifiable clinical notes from home video — to save nurses one third of their day.

Idea 4 — LifelongCare. A behaviour model that keeps learning as the patient changes across months and years — lands pile on your stated interest in life-long learning.

I have written a short pitch for each on the Research ideas page — happy to walk through whichever excites you.

1. What are the current priorities of the STARS team for the coming years?

2. Where do you see a place for my profile — healthcare + multi-modal fusion?

3. What role could fit best — postdoc, engineer, or researcher? What is the typical path?

4. Are there European projects starting soon where I could contribute quickly?

5. If we agree it is a good match, what would be the natural next step?