Computer Vision News - July 2022

BEST OF CVPR 26 CVPR Oral Presentation the images have to say about an object is not as powerful as trying to learn this attention scheme. What is Damien most proud of about this work? “ The way it’s coded ,” he responds. “It’s not something we pushed forward in the paper because it’s more of an engineering problem, but it required a great deal of effort, so I’m proud of what I’ve been able to build. It’s a hidden element, but it’s important so that you can do something in a reasonable time and in an efficient way.” Damien tells us the team is already discussing its next steps. They are looking at extending a paper that co-author Loic Landrieu worked on called Superpoint Graphs , designed to learn on point clouds at a large scale. “We’d like to keeppushing in this direction,” he affirms. As some objects can be seen by one image, some can be seen by several images, and some are not seen by anything, the team had to find a way to deal with this fluctuating information. To solve this, they used an a ttention mechanism . “We wanted the network to focus its attention across various inputs to combine their information,” he explains. “We evaluated our method on two large- scale data sets – one called S3DIS (Stanford 3D Indoor Spaces Dataset) , which has been used a lot as a benchmark in 3D semantic segmentation, and one called KITTI-360 , which is more outdoors – and obtained state-of-the-art results on both.” The team demonstrated much better results when extracting features from images than just colorizing the point clouds and hoping a 3D model would understand what the colors mean. Also, they proved that using a simple aggregation like taking the maximum or the average of what