A paper on CVPR 2017

One of the main events for the computer vision community is the annual conference Computer Vision and Pattern Recognition. This conference publications are the most highly cited not only in the field of computer vision, but also in computer science in general. Recently, the focus of the conference has shifted to deep learning, since deep convolutional neural networks have become the basis of most successful methods of computer vision.

In 2017, the conference was held from July 21 to 26 in Honolulu, USA. It was attended by about five thousand people (37% more than last year). In the main section of the conference 783 papers were presented (out of 2,620 submitted, acceptance rate 30%). In connection with the huge interest to computer vision technologies from the industry, this year the company's sponsorship contributions almost doubled: they amounted to 859 thousand US dollars. Among the sponsors was a large number of Asian companies, primarily Chinese. Also there was Russian sponsor Yandex.

In the main section of the conference, a poster report was presented by a research fellow of the laboratory Michael Figurnov. The work Spatially Adaptive Computation Time for Residual Networks is devoted to the problem of increasing the computational efficiency of convolutional neural networks. The article is written in collaboration with Google researchers, Carnegie Mellon University and Dmitry Vetrov. The motivation for this work was that the use of modern deep neural networks with hundreds of layers can be redundant, especially for images where a large area is occupied by an uninformative background. The article suggests a new deep architecture, which learns to adapt spatially its depth. Due to this, the pixels essential for the task being solved are processed by all available layers, and the non-informative pixels are only processed by several layers. Thus, the computational efficiency of the model is improved. The report aroused great interest of the conference visitors, in particular, representatives of the industry, for whom increasing the speed of models is the most important requirement for the products being developed.

Also at the conference, another faculty's employee, Artem Babenko, introduced paper Product Split Trees. It deals with search of the nearest neighbor for search engines. The solution of this task is required, for example, to implement search by pictures. The article proposes a new method, which shows the best ratio of speed and quality in comparison with previous works.