Work

A collaboration with Brno University of Technology (VUT), commissioned by the Ministry of the Interior for the Police of the Czech Republic’s Criminal Police Service. The tools automate the processing of images and video, focused on detecting and analysing human figures to support criminal investigations. The software is now in active use by the Police, and the project was awarded Excellent Research Results by the Ministry of the Interior in November 2025.

The project also produced an automated sensitive-content detection system, now deployed and used by the Czech Police. Its research outputs include ProbPose (CVPR 2025) and BBox-Mask-Pose (ICCV 2025). These papers address parts of the broader challenge of analysing people reliably in real-world imagery.

At Porsche Engineering Services in Prague (March 2020 – July 2022), I worked on production control-unit software for the Porsche Macan and software across the components of the High Power Charging platform, used by the Taycan and other electric vehicles. My work included C/C++, Plug & Charge and daily automated testing in Python. The two-page CV records more than 100,000 Macan vehicles and more than 1,000 charging units worldwide. These are deployment figures recorded in that CV, rather than live totals.

A project led by Matej Hoffmann studying infant behaviour to better understand human sensorimotor development. The contribution here is precise 2D pose estimation of babies in video, which supports analysing and interpreting infant movements and interactions — part of a larger effort to model human behaviour in its earliest stages. The face-anonymization work published as BLANKET (ICDL 2025) grew out of the same video data.

Outputs include BLANKET (ICDL 2025), on anonymizing infant faces, and private infant video data used to evaluate BBoxMaskPose.

A project led by Lukáš Picek (University of West Bohemia) identifying wild animals, particularly lynxes, from camera-trap images. The contribution here is 2D pose estimation of the animals, which improves identification accuracy — each lynx has a different fur texture on each side, and pose estimates help align the correct side for comparison. Part of a larger effort applying computer vision and AI to environmental protection.

SKV Revie saves the last 30 seconds of footage on demand, so coaches can pull up a clip and give feedback immediately after a play. The backend runs on a local mini-PC with a browser-accessible frontend, sustaining under 0.5 seconds of latency across at least 6 simultaneous 4K camera feeds. I built it as a web app so coaches could use it across operating systems and devices without installing anything.

Revie was used during the 2025/26 men’s play-down and the women’s playoffs. Since then, 10+ SKV coaches have used it daily during training. It remains closed-source for internal club use. If you’re a coach interested in using Revie, feel free to contact me.

An automated system for downloading video from Sokol Královské Vinohrady’s sports-hall cameras. The interface is available on the club’s staff-only website and allows downloading footage from any camera within a 30-day history. The backend connects a Python service, Microsoft Power Automate and Microsoft OneDrive, logs every run, and sends failure notifications so problems get caught quickly. It has handled over 750 download requests, with 1000+ videos downloaded through the system.

Tracking floorball players from multiple cameras, starting with a bachelor’s thesis on top-view person tracking and continuing through a master’s thesis on multi-camera identification and re-identification. The bachelor’s thesis system used a modified YOLO detector and a fine-tuned ResNet similarity metric, reaching 94% mAP detection and 90% MOTA tracking on a new sports-video dataset. The master’s thesis introduced an unsupervised, hard-negative-mining approach to training an identification network without labelled data, generating low-ID-switch tracklets suitable for labelling further training data. The applied interest in sports video continues in the Revie clip-retrieval system.

The related recording system serves more than 20 users around the clock and had downloaded more than 300 matches when the two-page CV was written. A server application with a desktop and mobile web interface supports real-time situation analysis during matches.

Research outputs: BSc thesis and MSc thesis.

A simple GUI tool for annotating 2D human pose in images — a lightweight, local alternative to CVAT and LabelStudio. Originally built to create the RePoGen dataset, it has since been reused for several other datasets. Open source, with contributions and feedback welcome.

Our group also uses PoseAnnotator for other structured keypoints, including facial landmarks, and for datasets published at CVPR and ICCV. It is available as an open-source Python package.