Computer Vision News - January 2022

13 DISCO: accurate Discrete Scale Convolutions we analyze signals with this DISCO approach. ” Ivan hopes the pair’s next paper will present their idea in an even clearer way so that everyone will be able to understand what it is about, why it is so important, and where it is applicable – if applicable. “ It’s always important to show there are limits, ” he points out. “ CVPR explicitly asks authors to write about that so it’s clear that if these limitations are against your approach, you can still read the paper, but it will not help you. The way we structure the research is very important. The next paper we write will have a focus on good presentation from the very start. ” He would like to see the model demonstrate a speed advantage over previous attempts because speed is so important for fields like autonomous vehicles. The price point is also key, and he would like to show it is affordable enough for people to equip their products, such as smartphones, with the technology. “ In this work, we pay attention to the computational complexity of other methods, ” Artem points out. “ In modern computer vision, in an age where we have these giant models, transformers which take weeks of training to converge, I think it’s very cool if we can impose some additional structure on our neural network. It allows us to train faster and to have some nice implementation, which means we also get faster inference. Speed and convergence of the networks is an important issue, and we try to tackle that in our work. ”