Constrained Deep Learning for Outlier Rejection


In this work we hope to combine the work from the robust optimization literature anddeep learning literature to create a general framework for robust matching. It can be usedto enhance performance in existing deep learning frameworks or to improve performancein robust learning frameworks. We leverage the aforementioned primal-dual trainingtechniques to learn more robust matching estimators. We formulate a Lagrangian primal-dual training framework for robust matching problems in asemi-supervised setting. Synthetic experiments have proved promising in a semi-supervised setting, and tests are underway for real-world data.

Camera Radar Fusion


From my previous work, I have focused on computer vision with multiple cameras to obtain geometric information such as pose. More recent trends, such as in the self-driving car industry, have focused on incorporating data from multiple complementory sensors. This is called sensor fusion, and while more complex has many advantages over using a single kind of sensor. Cameras and radars complement each other’s information quite well, but research on fusing the two has only recently started to gain interest. Using graph neural networks, we hope to be able to use data-driven methods to fuse the sensors more robustly in more settings.