Alexis Baudron Alexis Baudron

I research computer vision systems for cameras and sensors that need to operate in the real world. My work spans computational photography, synthetic data, and vision-language models for specialized image domains.

At Sony, I train vision and multimodal models for semiconductor imaging and yield optimization, combining sensor knowledge with machine learning and deploying agents that support high-volume manufacturing workflows.

I am deeply interested in

  • World models and sensing How richer learned representations can reshape cameras, sensors, and physical perception.
  • Robotics and embodied AI How world models, visual reasoning, and memory can help robots understand and operate in real environments.
  • Creative computational photography How optics, sensors, algorithms, and learning can be combined to build new ways of seeing.
  • VLM reasoning for specialized domains How vision-language models can be adapted to reason over technical visual data, from sensor evaluation to manufacturing.

work

2024 – Present

Sony Semiconductor Solutions

Senior AI Researcher · Pentas Vision

I lead a small team working on computer vision problems across sensors, manufacturing, and image quality. These days, we focus on VLMs and agents for semiconductor image reasoning and sensor evaluation. Previously we deployed diffusion-based synthetic data generation for edge vision models and optimization systems for industrial defect detection.

2021 – 2023

SenseBrain · SenseTime Research

Research Engineer · Computational photography & imaging

Novel sensors, ML-based artifact removal and image enhancement, and generative models for mobile imaging systems. Exploring the limits of smartphone cameras.

2016 – 2021

Northwestern University

B.S & M.S in Computer science · computational photography lab

My M.S. research was with Northwestern’s Computational Photography Lab, advised by Ollie Cossairt and Aggelos Katsaggelos, where I worked on event-camera 3D reconstruction. I also spent a lot of time at The Garage, Northwestern’s startup incubator, as both a builder and mentor.

talks & publications

Workshop
Visual Synthetic Data Generation for domain adaptation of edge vision models
Guest lecture @ University of Hawaii Manoa & DSS San Francisco
Talk
Talk
Paper
Paper

projects & builds

outside

Programmable Light Stage

A custom music stage built from scratch for Vitarave and built with friends, spanning wood fabrication, lighting hardware, DMX control, and software. Each panel is individually DMX-addressable, turning the backdrop into a programmable visual instrument that could be animated and synced live with the music.

Last shots show it in action at the festival.

Building the stage structure Lighting panels assembled on the stage Close-up of the individually DMX-addressable LED panels The stage lit up in action at the festival The stage in action during the show

3D with event-cameras

E3D was my M.S. research project on recovering 3D shape directly from event-camera data. Inspired by earlier event-camera reconstruction work from our lab, we asked whether the sparse edges and contours produced by an event camera could be used to reconstruct dense object geometry.

The pipeline first converts events into silhouettes and camera poses, then uses differentiable rendering to optimize a dense 3D mesh whose projected shape matches the event-derived observations. The first figure shows the full pipeline, followed by reconstruction results on synthetic and real event sequences.

E3D was one of the first papers to target dense mesh recovery from monocular event cameras, and the area has since expanded into neural rendering, NeRF-style reconstruction, and event-based 3D Gaussian Splatting.

E3D pipeline: from events to dense 3D mesh E3D reconstruction results on synthetic and real event sequences

Container tracker

Freight container tracking system I built as a prototype for a client at OneTrack. The unit was fully power-autonomous, drawing energy from a front-mounted solar panel and storing it in a car battery throughout the day. It was deployed outdoors and survived multiple Chicago snowstorms.

The system used two cameras working together. A wide-FOV camera ran an SSD-based truck detector I trained to track freight containers across frames and determine when a container entered the capture zone. Once triggered, a narrow-FOV camera captured a burst of higher-resolution images, which were processed by a probabilistic container identification system I built.

Container tracker Container tracker Container tracker Container tracker Container tracker

Intelligent Flying Machines

Final drone platform I built during my time at Intelligent Flying Machines (before we became OneTrack.ai). The frame was fully CNC-machined from carbon fiber and used no 3D-printed structural parts. This design was the result of a year-long iteration cycle across multiple drone platforms, ultimately reducing weight by 20% and increasing flight time by 17%. A few of the key mechanical details included a custom vibration-isolation plate for the IMU and a novel docking mechanism designed for more reliable field operation.

Also here's a picture of me with the custom docking mechanism I built for one of the earlier drones we had.

Prototype warehouse drone Drone build Drone hardware Me with the custom docking mechanism

🚲 I love biking!

I used to race triathlon at Northwestern. These days I mostly just ride bikes early mornings, long climbs, coffee stops, and the occasional sunrise that makes it all worth it.

Out on the bike Cycling

🍳 chefing it up!

I like cooking things and bringing people together! pictures: seared scallops, crispy rice, edible flowers, whole pig in a caja china, focaccia

Seared scallops and a caramelized terrine in herb oil Crispy rice bites topped with scallop and edible flowers Roasting a whole pig in a Caja China with friends foccaccia