Computer Vision News - July 2021

Fabio Tosi / Yiyi Liao 31 Best of CVPR 2021 The overall goal for this work is to build a stereo matching algorithm that can work at a very high resolution and predict sharp and precise object boundaries. The team have a sensor in their lab that captures at 12 Mpx resolution, which they want the algorithm to work with, and an algorithm capable of providing precise 3D geometry was desirable. The team felt uncertainty was important. Typically, when considering uncertainty, a single-modal distribution is used. Fabio and Yiyi thought a multi- modal distribution, in particular a bimodal solution, would be a better alternative for stereo matching. This allows recovery of the sharp object boundaries. They demonstrated the flexibility of their technique by improving the performance of a variety of stereo backbones.