%0 Journal Article %1 laupheimer2020importance %A Laupheimer, Dominik %A Shams Eddin, Mohamad Hakam %A Haala, Norbert %D 2020 %E Kersten, Thomas P. %J Wissenschaftlich-Technische Jahrestagung der DGPF in Stuttgart – Publikationen der DGPF %K conference review from:markusenglich %P 205--218 %T The Importance of Radiometric Feature Quality for Semantic Mesh Segmentation %U https://www.dgpf.de/src/tagung/jt2020/proceedings/proceedings/papers/27_DGPF2020_Laupheimer_et_al.pdf %V 29 %X We propose a pipeline for the semantic segmentation of textured meshes in urban scenes as generated from imagery and LiDAR data. Key idea is to represent the mesh as a set of face centroids (COG cloud). This enables the comparison of various point-based classifiers of varying learning abilities. Fine-tuned PointNet++ showed the best results due to hierarchical feature learning. One of the main differences between meshes and point clouds is the availability of high-resolution texture. Hence, we evaluate the importance of radiometric feature quality as a proxy for texture importance. Color information increases performance by at least 5 % (mIoU) for the used data. We achieved to double the performance gain by improving the radiometric feature quality, i.e. utilizing color information of the entire face. Our study shows that texture is beneficial for non-uniform dense and non-balanced data sets. However, it also shows the inherent limitations of textural features like occlusions, absence of imagery, and the quality of the geometric reconstruction

@article{laupheimer2020importance, abstract = {We propose a pipeline for the semantic segmentation of textured meshes in urban scenes as generated from imagery and LiDAR data. Key idea is to represent the mesh as a set of face centroids (COG cloud). This enables the comparison of various point-based classifiers of varying learning abilities. Fine-tuned PointNet++ showed the best results due to hierarchical feature learning. One of the main differences between meshes and point clouds is the availability of high-resolution texture. Hence, we evaluate the importance of radiometric feature quality as a proxy for texture importance. Color information increases performance by at least 5 % (mIoU) for the used data. We achieved to double the performance gain by improving the radiometric feature quality, i.e. utilizing color information of the entire face. Our study shows that texture is beneficial for non-uniform dense and non-balanced data sets. However, it also shows the inherent limitations of textural features like occlusions, absence of imagery, and the quality of the geometric reconstruction}, added-at = {2021-12-01T15:45:57.000+0100}, author = {Laupheimer, Dominik and Shams Eddin, Mohamad Hakam and Haala, Norbert}, biburl = {https://puma.ub.uni-stuttgart.de/bibtex/2d44c78da6c13bd7da6e0c6404fe64cb4/ifp}, editor = {Kersten, Thomas P.}, interhash = {bacce802ad8359477bc1c47b63639971}, intrahash = {d44c78da6c13bd7da6e0c6404fe64cb4}, issn = {0942-2870}, journal = {Wissenschaftlich-Technische Jahrestagung der DGPF in Stuttgart – Publikationen der DGPF}, keywords = {conference review from:markusenglich}, pages = {205--218}, timestamp = {2021-12-01T14:45:57.000+0100}, title = {The Importance of Radiometric Feature Quality for Semantic Mesh Segmentation}, url = {https://www.dgpf.de/src/tagung/jt2020/proceedings/proceedings/papers/27_DGPF2020_Laupheimer_et_al.pdf}, volume = 29, year = 2020 }