Ondrej Chum
Associate Professor
I am leading a team within the Visual Recognition Group at the Department of Cybernetics, Faculty of Electrical Engineering, Czech Technical University in Prague. I received the MSc degree in computer science from Charles University, Prague, in 2001 and the PhD degree from the Czech Technical University in Prague, in 2005. From 2006 to 2007, I was a postdoctoral researcher at the Visual Geometry Group, University of Oxford, United Kingdom. My research interests include large-scale image and particular object retrieval, object recognition, and robust estimation of geometric models. I am a member of Image and Vision Computing editorial board, and I have served in various roles at major international conferences (e.g., ICCV, ECCV, CVPR, and BMVC). I co-organize Computer Vision and Sports Summers School in Prague. I was the recipient of the Best Paper Prize at the British Machine Vision Conference in 2002. I was awarded the 2012 Outstanding Young Researcher in Image & Vision Computing runner up for researchers within seven years of their PhD. [. . .]


Our paper Mukundan, Tolias, Chum : Multiple-Kernel Local-Patch Descriptor, BMVC 2017 was awarded the Best Science Paper Honorable Mention.

Our paper Philbin, Chum, Isard,Sivic, Zisserman : Object Retrieval with Large Vocabularies and Fast Spatial Matching, CVPR 2007 was awarded the Longuet-Higgins Prize.


Ph.D. Students



Karlovo namesti 13, 121 35 Praha 2, Czech Republic
Office: G3 (building G, room 3), directions
Tel: +420 2 2435 7282

Projects and code

Unsupervised fine-tuning of CNN for retrieval
Training data, fine-tuned deep neural networks and matlab evaluation code to reproduce the results of our ECCV 2016 paper. See the project page.
Diffusion on Region Manifolds
Diffusion paper on arXiv and the project page.
Low dimensional feature maps
Feature maps construction for kernels with low-dimensional compact interval input cast as a convex optimization problem. See our ICCV 2015 paper for details, Matlab code is available.
Fast and reliable two-view geometry estimation from our BMVC 12 paper , software page, and data. Results any RANSAC paper should compare to.
USAC: Universal RANSAC
A modular framework implementing a number of RANSAC extensions. Our PAMI 2013 paper and code and data zip archive are available.
Homography from two correspondences
Homography estimation from a correspondence of at least two elliptical featires: ICPR 2012 paper and Matlab code .
Dataset: Alps100K
Alps100K dataset contains 98,136 annotated (GPS coordinates, elevation, EXIF if available) outdoor images from mountain environments, from our BMVC 2015 paper. See the project page.