Computer Vision News - December 2019

Kwang Moo Yi 17 Best of ICCV 2019 challenge is simple because the core thing is just representing images as local images and the rest of the pipeline is the same as other existing methods. This is really a case of a simple idea, well executed, that does make a difference. He adds that he is proud that they have revived something that was a direction that people were trying to take in the past. He tends to do a lot of things drawing inspiration from past works and believes log polar is one of those. There’s also something called Daisy that was developed by CVLAB at EPFL. It’s similar but there are some differences: Daisy used to create statistical features from normal typical images and then did the aggregation in log-polar domain, whereas in this case, they do it from the beginning because deep networks allow this. Kwang Moo is not actually the main author of the paper. That is Patrick Ebel , who has moved on to a different institution now. Also, Anastasiia Mishchuk , the second main author, is unavailable to talk to us today, so it falls on the third guy! He modestly tells us that they deserve more credit than him, but we can attest to the fact that the poster session had a full house on Tuesday, so it all worked out well in the end. He points out: “One of the things that was really good for our poster was that we didn’t have to take long to explain because it’s a single idea . It’s really nice that we’re going to continue investigating this because it fits perfectly into what we’ve been doing for three or four years now: learning local features. This will be an opener for another chapter for this.” Finally, Kwang Moo says the principal supervisor of the work, Eduard Trulls , has been a great partner for him in this journey of local features. Together with Pascal Fua , they are determined to continue working on this, even if they are all in different places.