Image Matching Challenge 2020

Local features have played a key role in a wide range of computer vision applications throughout the past 20 years. Despite the drastic advancements in many fields after the deep learning revolution, end-to-end solutions for 3D reconstruction under challenging conditions remain elusive, and solutions based on traditional, hand-crafted methods such as SIFT (1999) and RANSAC (1981) may outperform the state of the art in machine learning.

Historically, local features have been evaluated on small datasets using proxy metrics, such as patch retrieval accuracy for descriptors. This may not necessarily translate to downstream applications, misrepresenting their actual performance and hindering the development of new techniques. There is a clear need for large-scale, challenging benchmarks to both train and evaluate new strategies for image matching. Towards this end, we held in 2019 an open challenge in image matching across wide baselines. In contrast to existing benchmarks, we only consider integrated solutions and our primary metric measures the accuracy of the reconstructed poses — the final goal.

We are pleased to introduce a second, much improved edition of this challenge. It is co-located with the CVPR 2020 workshop on Image Matching: Local Features and Beyond. Winners and selected participants will be invited to give talks at the conference, as last year. Prizes are available, thanks to our sponsors.

** PUBLIC SERVICE ANNOUNCEMENT **: While the 2020 challenge is over, we plan to take submissions to current version of the benchmark until the next version is announced. Some of our resources are being transferred from UVIC to UBC, so please let us know if you find any issues.


Key Dates


Eduard Trulls Google
Yuhe Jin UBC
Kwang Yi UBC
Dmytro Mishkin CTU Prague
Jiri Matas CTU Prague
Pascal Fua EPFL