9 DAILY WACV Monday Filippo D’Addeo - Lorenzo Cipelli Lorenzo explains that all the methods trying to address the calibration between sensors focus on initial classification and then iterative refinement of the alignment of the two modalities. These gave mediocre results and very slow algorithms. Subsequent methods focused on trying to get features between the two modalities as close as possible. “In our method,” he adds “we try to follow this path but in a different way, meaning that we use also geometrical information which is inside or let's say intrinsic into a bird eye view representation, meaning an image of the point cloud from above. So we decided to make features closer between the two modalities and also leverage the geometrical aspect of bird eye view representations.” Then Lorenzo tell us of the challenges that they encountered on the way: “What I remember from our days in the office was that it is really a delicate task, meaning you change something and it could be that you don't get the outcome that you expected and it is not consistent many times.” They solved this with a lot of attention to the details. In some cases, Filippo adds, that meant writing the algorithms on a whiteboard, check for everything, visualize and see if they were right or wrong. They declare that, fortunately, they achieved robust results, meaning that the error is not deviating a lot from the mean of their results. They decided to start their work not from other calibration works but take instead inspiration from another task, which is the object detection task. Many people are more familiar with detection instead of calibration. The idea was to somehow adapt something made for detection, 3D detection, to the calibration task. They did prove in their work that this is possible, obtaining also better results compared to previous calibration methods.
RkJQdWJsaXNoZXIy NTc3NzU=