Idea: Because there is one free speaker slot during this week's meeting, everyone can present a highlight from CVPR that they think is interesting to the group. If you want to highlight a paper/workshop add it below and say a couple of words during the meeting.

SuperGlue is a graph neural network that simultaneously performs context aggregation, matching, and filtering of local features for wide-baseline pose estimation. It won three CVPR challenges and more information can be found on the project page or in the paper.

A real-time segmentation approach that combines classical ideas from snakes and uses geometric deep learning to refine a contour. More information can be found in the paper.

While we often use vision and inertial measurements to complement each other, there is a line of work that uses only an IMU, but incorporates prior knowledge about the type of motion. This work learns typical motion patterns using deep learning. workshop talk project page

Tutorial on how to write a good paper, review and rebuttal with some background info on how the whole process works. Had some great insights and good overview in general. tutorial

Task dependent image translation from simulated to real world images using an RL scene consistency loss that ensures the source and the generated images are invariant with respect to their corresponding Q-values. paper

A novel way to solve Localization and Mapping problem distributedly is proposed. The factor graph representation of bundle adjustment is adopted and an existing graph algorithm coined "Bielf Propagation" is proposed to solve the graph inference problem. They show >20x speedup of the proposed method running on IPU link, than traditional bundle adjustment on CPU. paper Teaser