SUM Parts: Benchmarking Part-Level Semantic Segmentation of Urban Meshes

Weixiao Gao, Liangliang Nan, Hugo Ledoux
Delft University of Technology
CVPR 2025
Paper Code (Coming Soon) Data (Coming Soon) Poster Video arXiv

Overview

Dataset Overview

SUM Parts provides part-level semantic segmentation of urban textured meshes, covering 2.5km² with 21 classes. From left to right: textured mesh, face-based and texture-based annotations. Classes include unclassified unclassified, terrain terrain, high vegetation high vegetation, water water, car car, boat boat, wall wall, roof surface roof surface, facade surface facade surface, chimney chimney, dormer dormer, balcony balcony, roof installation roof installation, window window, door door, low vegetation low vegetation, impervious surface impervious surface, road road, road marking road marking, cycle lane cycle lane, and sidewalk sidewalk.

Video Presentation

Annotation

Annotation

Our annotation aims to achieve precise semantic labeling with significantly improved efficiency for urban meshes. Our tool features two main modules for part-level semantic annotation: face-based annotation for triangle faces and texture-based annotation for texture pixels. We enhance the efficiency of both modules by incorporating interactive selection and template-matching strategies. We invited five individuals with experience in remote sensing to manually annotate the dataset using our tool: two focused on face-based annotation, two on texture pixel-based annotation, and one reviewed and corrected the annotations. The entire annotation process took approximately 640 hours in total.

Benchmark Datasets

Benchmarks

We defined two label types: face (12 labels, excluding `unclassified`) and pixel (19 labels, excluding `terrain` and `unclassified`). For face labels, we evaluated four mesh point cloud sampling strategies; random/Poisson-disk sampling matched superpixel texture sample size, while face-centered sampling matched mesh face count. For pixel labels, we tested three sampling methods: random, Poisson-disk, and superpixel texture sampling. We also evaluated state-of-the-art 3D semantic segmentation methods, including mesh-based (RF-MRF, SUM-RF, PSSNet) and point cloud-based (PointNet, PointNet++, SPG, SparseConvUnet, RandLA-Net, KPConv, PointNext, PointTransV3, PointVector) approaches.

BibTeX


            @InProceedings{Gao_2025_CVPR,
            author    = {Gao, Weixiao and Nan, Liangliang and Ledoux, Hugo},
            title     = {SUM Parts: Benchmarking Part-Level Semantic Segmentation of Urban Meshes},
            booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
            month     = {June},
            year      = {2025},
            pages     = {24474-24484}
            }
        
            
            @article{sum2021,
            author = {Weixiao Gao and Liangliang Nan and Bas Boom and Hugo Ledoux},
            title = {SUM: A Benchmark Dataset of Semantic Urban Meshes},
            journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
            volume = {179},
            pages = {108-120},
            year={2021},
            issn = {0924-2716},
            doi = {10.1016/j.isprsjprs.2021.07.008},
            url = {https://www.sciencedirect.com/science/article/pii/S0924271621001854},
            }