New preprint: PEEK — frame selection for efficient video captioning
We open sourced the code and model weights on HuggingFace, and released a live demo you can try in the browser.
Industrial PhD / Computer Vision / Multimodal Learning
Killian Steunou
Building efficient visual systems for the multimodal world.
I am an industrial PhD student at Institut Polytechnique de Paris and Moments Lab, working on efficient omni-modal learning for generalized video understanding.
PhD
Industry and academia, since November 2025
Vision
Video understanding, tracking, representation learning
Open
Reproducible research and transparent engineering
Current thesis
Efficient omni-modal learning for generalized video understanding.
Research themes
Efficiency, multimodal learning, tracking, and deployment-aware model design.
Working style
Scientific rigor, clear writing, reproducible pipelines, and pragmatic engineering.
News
About
A research practice shaped by efficiency, clarity, and deployment constraints.
I am currently pursuing an industrial PhD in machine learning at Institut Polytechnique de Paris and Moments Lab. The work sits at the intersection of modern multimodal models and the practical limits that define whether they remain useful in the real world.
I am especially drawn to problems in computer vision and deep learning because visual perception feels foundational to intelligence, both human and artificial. I like models that can scale, adapt, and still remain legible enough to improve.
Current focus
Efficient omni-modal learning for generalized video understanding, with an emphasis on scalable training and practical inference.
Research values
Open-source practices, transparent experimentation, and reproducible pipelines that other researchers can meaningfully build on.
LinkedIn
GitHub
Scholar
CV
Experience
Industrial research, product-minded experimentation, and model-building in production settings.
Machine Learning PhD Student
Moments Lab
- Researching efficient omni-modal learning strategies for generalized video understanding.
- Working on problems where representation quality, inference cost, and deployment realism all matter.
Deep Learning Research Engineer Intern
Idemia
- Explored how SAM 2 can strengthen end-to-end multi-object tracking systems.
- Built segmentation-aware and proposal-aware tracking variants around MOTIP-style models.
AI Research Intern
Collecte Localisation Satellites
- Developed tooling to fine-tune foundation models on remote sensing datasets.
- Benchmarked segmentation performance and summarized the rapidly changing literature.
Machine Learning Engineer Intern
JoliBrain
- Integrated ControlNet-style controls and zero-shot detection models into joliGEN.
- Helped ship documentation and product-facing material around the tool.
Software Developer Intern
French Ministry of Agriculture
- Created agreste, an R package and R Shiny workflow for statistical publication automation.
- Handled the project end to end, from requirements gathering to delivery.
Education
A mathematics and statistics foundation, refined through machine learning and vision research.
Machine Learning PhD
Institut Polytechnique de Paris
- Research topic: efficient omni-modal learning for generalized video understanding.
Master 2 Mathématiques, Vision, Apprentissage
ENS Paris-Saclay
- Optimal transport, convex optimization, deep learning, graphical models, and generative models.
Master 1 Applied Mathematics and Statistics
Toulouse School of Economics
- Econometrics, probability, optimization for ML, time series, and data science in Python.
Gap Year
University of Copenhagen
- Natural language processing, blockchain business development, energy economics, and tax policy.
Double Bachelor: Applied Mathematics and Economics
Toulouse School of Economics
- Linear algebra, analysis, statistics, econometrics, optimization, programming, and economics.
Publications
Preprint and Accepted Papers
2026
arXiv Preprint
PEEK: Picking Essential frames via Efficient Knowledge distillation
We propose PEEK, a training-efficient frame selection method that distills knowledge from vision-language teachers to identify the most informative frames for video captioning — achieving competitive performance while significantly reducing the number of processed frames.
2025
arXiv Preprint
Sparse Representations Improve Adversarial Robustness of Neural Network Classifiers
We show, theoretically and empirically, that SPCA-based classifiers are more robust than PCA-based alternatives under adversarial attack, providing a new perspective on linear dimensionality reduction as a defense mechanism.
Projects
Tools, experiments, reproductions, and interfaces that turn research ideas into usable artifacts.
Contribution to joliGEN
I integrated ControlNet-inspired edge controls and SAM-based masking into joliGEN, then helped improve its documentation.
Video Object Detection
An implementation of Owl-ViT for zero-shot object detection in videos using natural-language prompts.
An End-to-End Transformer Model for 3D Object Detection
We reproduced 3DETR on SUN RGB-D and explored how a lean transformer detector behaves when extended with RGB information.
Wikipedia Graph
A French-language concept graph built by scraping linked Wikipedia topics and exporting the result for graph exploration.
Word Ladder Generator
A French word-ladder generator that computes the shortest sequence of one-letter edits between two words.
Writing
Long-form notes that turn research trends and evaluation tools into something easier to reason about.
Efficiency Follows Capability: A Decade of Video Understanding Research Trends
An analysis of how efficiency moved from the margins to the center of video understanding research from 2015 to 2025.
NLP Metrics for Image & Video Captioning: A Visual Guide
A worked visual walkthrough of n-grams, TF-IDF, BLEU, ROUGE-L, METEOR, and CIDEr for captioning research.
Demos
Interactive spaces for testing models and interfaces beyond the paper.
PEEK — Frame Selection for Video Captioning
Select the most informative frames from a video, distilled from vision-language teachers.
Audio Visual Transcription
Whisper-powered subtitling for audio and video with a fast demo workflow.
Video Background Removal
Interactive segmentation-based background removal for video.
Contact
If you care about efficient vision systems, robust evaluation, or research that has to ship, we should talk.
I am always interested in thoughtful conversations around machine learning research, multimodal systems, visual understanding, and the engineering decisions that make models usable outside the lab.
Video understanding
Computer vision
Multimodal ML
Research collaborations