Researchers from the Czech Institute of Informatics, Robotics and Cybernetics (CIIRC CTU), including Torsten Sattler, Patrik Beliansky, Tomas Pajdla, Josef Sivic, Martin Cifka and Anna Sarova Mikestikova presented the results of their research at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), held in Denver, Colorado, USA, from June 3 to 7, 2026. One of the world’s leading conferences on computer vision, CVPR 2026 attracted a record number of submissions, with 4,072 papers accepted out of 16,092 submissions.
Two Papers Presented at the Image Matching Workshop
Two of the four papers were presented at the Image Matching Workshop, co-organized by Prof. Jiri Matas and Dmytro Mishkin from the Faculty of Electrical Engineering (FEL) at CTU.
The first paper, Can We Make NeRF-Based Visual Localization Privacy-Preserving? (Maxime Pietrantoni, Martin Humenberger, Torsten Sattler, Gabriela Csurka), considers the problem of privacy-preserving visual localization. Visual localization is the problem of determining from which position and orientation a given query image was taken. Privacy-preserving visual localization aims to localize images while revealing as little privacy-related content as possible, both in the query images and in the 3D scene representation used for localization. The authors propose a scene representation based on neural radiance fields (NeRF) that uses image segmentations rather than raw images to represent the query images and to train the scene representation. The paper was presented by Maxime Pietrantoni.
The second paper, Privacy-Preserving Structureless Visual Localization via Image Obfuscation (Vojtech Panek, Patrik Beliansky, Zuzana Kukelova, Torsten Sattler), also focuses on privacy-preserving visual localization. The paper shows how a simple visual localization pipeline can be made privacy-preserving by obfuscating both the query images and the scene representation, for example by using semantic segmentations instead of the original images. The work was carried out in collaboration with Zuzana Kukelova from FEL CTU. The paper was presented by Patrik Beliansky.
Both papers were selected for oral presentations.
Poster Presented in the Findings Track
The paper Active Exploration for Sparse Visual Localization (Johanna Lidholm, Ludvig Dillén, Zuzana Kukelova, Torsten Sattler, Viktor Larsson) was presented as a poster in the newly introduced Findings Track. The work explores how to make a standard visual localization pipeline more efficient and robust through local exploration of the scene. The research was conducted in collaboration with Zuzana Kukelova from FEL CTU and Viktor Larsson’s group at Lund University in Sweden.
Paper Selected as a CVPR Highlight
The paper Simple but Effective Triplet-Based Compression Strategies for Compact Visual Localization (Torsten Sattler, Zuzana Kukelova) was presented as a poster in the main conference.
The paper considers how to make scene representations used for visual localization more compact, reducing memory requirements while maintaining accurate localization. To achieve this, the authors propose a set of simple compression strategies that are almost trivial to implement yet reach or surpass the current state of the art.
The paper, a collaboration with Zuzana Kukelova from FEL CTU, was selected as a CVPR Highlight. Only 578 papers, representing 14.1% of the 4,072 accepted papers, received this distinction.
Papers from CIIRC CTU in the main conference track
Other CIIRC researchers also contributed to the CVPR program.
Tomas Pajdla co-authored the paper Minimal Constraint Relaxation for Multiview Autocalibration (Norio Kosaka, Timothy Duff, Tomas Pajdla), which addresses multiview camera autocalibration.
Josef Sivic co-authored AlignPose: Generalizable 6D Pose Estimation via Multi-view Feature-metric Alignment (Anna Šárová Mikeštíková, Médéric Fourmy, Martin Cifka, Josef Sivic, Vladimir Petřík), which focuses on generalizable 6D object pose estimation through multi-view feature-metric alignment.
