Computer Vision News - February 2023

19 Larissa Triess critical scenarios. Therefore, Larissa investigated three different approaches on how to achieve more robust performance in the application domain without any additional annotation effort: 1) It is common to insert additional objects into the scenes at training time. This is especially helpful for detecting under-represented dynamic objects. Larissa’s work focused on the generation of such 3D objects [ Triess2022AISTATS ] . 2) Domain mapping is a type of domain adaptation where data from a source domain is transformed into a target domain. This is achieved with generative models that synthesize realistic high-resolution LiDAR data [ Triess2019IV ] . A perception model can then be trained on the transformed data and be applied in the application domain without being exposed to a domain shift. In her work, Larissa proposes a quantitative metric to judge the quality of the transformed LiDAR data [ TriessGCPR2022 ] . Figure 2 shows the concept and example output of the metric that provides local realism scores for the point clouds. In a detailed analysis, Larissa shows how the data quality and the perception performance are correlated [ Triess2022IJCV ] . 3) For learning domain-generalized features, Larissa uses a method that simultaneously learns the semantics and the geometry of the scene. Since the geometry can be learned in a self-supervised fashion, it is advisable to learn them jointly for the source and target domains. The domain-generalized features then help to capture the semantics of the target domain. To summarize, Larissa’s dissertation presents a comprehensive framework for domain adaptationandsemantic segmentationof LiDARpoint clouds inthecontextof autonomous driving. For updates on her future work, follow Larissa on LinkedIn . Figure 2: The realism measure has a tripartite understanding of the 3D-world (middle). It estimates the realism of local regions within the point cloud. Both example images are real-world scenes (green) with augmentations from a simulation framework (blue) and point distortions (red).