Available for contract & remote work

Valerijan Matvejev

Computer Vision Engineer 3D/2D Pipelines Monocular 3D Human Motion Reconstruction PhD Researcher

Scroll

01About

Turning ordinary video into precise 3D understanding.

I'm a computer vision engineer and PhD researcher. I build 3D/2D vision pipelines that turn everyday camera footage into accurate, physically grounded 3D data.

At SportFX I work on monocular 3D human motion reconstruction: recovering metric body motion from a single camera and validating it against marker-based ground truth. In parallel I'm a PhD researcher at the Laboratory for Data Analysis and Applied AI at ETF, University of Belgrade, working on 3D vision pipelines and AI security, in research collaboration with École Polytechnique in Paris.

Earlier work spans eye-tracking-based memorability prediction and satellite image segmentation. I also enjoy explaining AI to the public: 10+ talks and media appearances so far.

Now
Computer Vision Engineer · SportFX2026 – present · Remote
PhD
Computer Science & AI · ETF, University of Belgrade2024 – present
Lab
Laboratory for Data Analysis and Applied AI · ETFResearcher
Paris
Research collaboration · École PolytechniqueFrench Government Scholarship · renewed October 2026 · ongoing
Before
Junior Computer Vision Researcher · BioSense Institute2024 – 2025 · Novi Sad
M.Sc.
Visual Computing · Nantes Université (Polytech Nantes)2023 – 2024 · University of Nantes scholarship
B.Sc.
Software Engineering · ETF, University of Belgrade2018 – 2023 · GPA 9.48 / 10
Lang.
English C2 · French B2Cambridge CPE · DELF B2
  • PyTorch
  • TensorFlow
  • OpenCV
  • NumPy
  • Monocular depth
  • 3D human mesh recovery
  • Pose estimation
  • Camera calibration
  • LiDAR / point clouds
  • Signal filtering
  • Blender
  • Python
  • C/C++

Degrees & distinctions

M.Sc. Visual ComputingNantes Université · 2024
B.Sc. Software EngineeringUniversity of Belgrade, ETF · 2023 · GPA 9.48
Dositeja AwardFund for Young Talents of Serbia · 2022

02Featured work / SportFX

From a single camera to metric 3D human motion.

As a Computer Vision Engineer at SportFX, I work on a pipeline that reconstructs full-body 3D human motion from ordinary monocular video. It recovers metric body motion from a single camera, then makes it physically consistent so the result can be measured, compared and trusted.

SportFX product reel · © SportFX

↳Results in production

  • Zerofoot–ground penetration
  • −83%foot skating
  • −99%acceleration jitter
  • 17.6 → 2.1cmcenter-of-mass 3D trajectory error vs. Qualisys ground truth

A physics-aware grounding optimization, validated against Qualisys marker-based motion capture and cross-checked with ARKit IMU data.

  1. 01

    Single-camera input

    One ordinary RGB video. No markers, no suits, no studio rig.

  2. 02

    Metric 3D reconstruction

    Full-body motion at real-world scale, built on state-of-the-art monocular depth models and multi-modal sensor fusion.

  3. 03

    Physics-aware grounding

    An optimization that keeps feet on the ground: no penetration, far less skating and jitter.

  4. 04

    Ground-truth validated

    Benchmarked against Qualisys marker-based motion capture and ARKit IMU data.

03Visual attention / Master's research

Where people look, and what they remember.

In an eye-tracking study with real participants, carried out at the ETF Innovation Center for my master's thesis, I turned fixations into spatial attention maps and trained a machine learning model to predict whether an image will be remembered.

92% accuracy predicting
image memorability

M.Sc. Visual ComputingPolytech Nantes

IEEE TELFOR 2025 paper
  • Eye-tracking
  • Fixation maps
  • Visual attention
  • Memorability
  • TensorFlow

04Remote sensing / BioSense Institute

Reading the Earth from orbit.

As a Junior Computer Vision Researcher at BioSense, the European Institute of Excellence, I developed segmentation models for large-scale satellite imagery. That included zero-shot delineation of agricultural parcels with Segment Anything (SAM), and cross-sensor analysis between PlanetScope (3 m) and Sentinel-2 (10 m).

93% segmentation accuracy
on Sentinel-2 imagery

Junior CV ResearcherBioSense Institute2024 – 2025

  • Sentinel-2
  • PlanetScope
  • Segment Anything (SAM)
  • Semantic segmentation
  • PyTorch

05AI security / École Polytechnique

Data Taggants: proving who owns a dataset.

Dataset ownership verification through harmless, targeted data poisoning. Data taggants embed verifiable ownership signals into a dataset, so models trained on it can later be identified, without degrading performance or visibly altering the data. I'm extending the method to segmentation models.

  • VerifiableEmbedded signals reveal when a model was trained on the protected data.
  • HarmlessNo degradation in model performance.
  • InvisibleNo visual alteration of the data itself.

Research collaboration with Prof. El Mahdi El MhamdiÉcole Polytechnique, ParisRenewed Oct 2026 · ongoing

data_taggants_segmentation

06Research home / ETF Belgrade

Laboratory for Data Analysis & Applied AI.

University of Belgrade · School of Electrical Engineering (ETF)

My PhD research home. The lab builds software systems on machine learning and deep learning, works on natural language processing for Serbian and on computer vision, and runs data analysis projects in education, medicine and transport.

  • 1971Roots in ETF's first Data Processing Laboratory. Reopened in 2016 as its successor.
  • DIHThe Belgrade Data Innovation Hub was formed on the lab's foundations.
  • 2019Members co-authored Serbia's National AI Strategy and helped form the National Institute for AI (2021).
Lab website · data.etf.bg.ac.rs
Flagship project · Science Fund of the Republic of Serbia (PROMIS) STOP

AI-driven hate speech detection for the Serbian language.

Software for Text Offences Prevention: a new dataset of Serbian short texts, Serbian NLP models and a software system that detects hate speech. Presented on RTS and at international conferences. Principal investigator: Prof. Dražen Drašković.

stop.etf.bg.ac.rs

The people I work with

07Publications & talks

Papers, talks and media.

  1. 2026

    A Robust Postprocessing Grounding Framework for Monocular 3D Motion Recovery: Benchmark on MLB Baseball Footage

    V. Matvejev, D. Drašković

    5th Serbian International Conference on Applied Artificial Intelligence (SICAAI) · Kragujevac, Serbia

  2. 2025

    Web Application for Predicting Image Memorability Using a Machine Learning Model

    V. Matvejev, D. Drašković, B. Nikolić, D. Bojić

    IEEE · 33rd Telecommunications Forum (TELFOR 2025) · Belgrade, Serbia

    DOI 10.1109/TELFOR67910.2025.11314318
  3. 2025

    Prediction of Image Memorability Using Machine Learning Models

    V. Matvejev, D. Drašković

    31st ICT Conference YU INFO 2025 · Kopaonik, Serbia · in Serbian

    PDF
  4. 2024

    Development of a Web System with an Automated Question Generator Based on Large Language Models

    L. Hrvačević, J. Cincović, A. Milaković, V. Jocović, V. Matvejev, D. Drašković

    IEEE · 11th International Conference IcETRAN 2024 · Niš, Serbia

    DOI 10.1109/IcETRAN62308.2024.10645106

09Beyond the lab

Off the clock, still in motion.

I reconstruct human motion for a living, and produce plenty of it in my free time: long runs, mountain passes on a bike, and talks about AI for anyone who'll listen.

3D mesh + skeleton Photo
Monocular 3D, on myselfOne photo in, body mesh and skeleton out. Drag to reveal.
CyclingSedlo pass · Durmitor, Montenegro
Off dutySame curiosity, fewer GPUs
10+ talks & media appearances Explaining AI to the public, on stage and in the media. All talks
Running21 km along the Danube · fastest 20K yet

10Contact

Have a vision problem? Let's talk.

Open to contract and remote work in 3D/2D vision pipelines.