Research map / Detection and segmentation / Open-set and few-shot detection
SUM Parts: Benchmarking Part-Level Semantic Segmentation of Urban Meshes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces SUM Parts, a 2.5 km² part-level semantic segmentation dataset for urban textured meshes with 21 classes, plus an interactive face/texture annotation tool and benchmarks of 3D segmentation and interactive annotation methods.
Problem
Urban scene semantic segmentation has focused on images and point clouds, leaving textured meshes and part-level (functional component) labels largely unexplored, and mesh annotation is slow and sparse in tooling.
Approach
Annotate urban textured meshes at the face and texture-pixel level using an interactive tool combining graph-cut protrusion extraction, 3D/2D template matching, and superpixel-based texture annotation, then benchmark point-cloud and mesh segmentation methods on sampled points from meshes.
Key results
- PointVector achieves the best face-labeling result with 70.0% mIoU and 80.7% mAcc over 12 face classes.
- On pixel labeling (19 classes), PointVector with superpixel texture sampling reaches 47.9% mIoU, well above other point-cloud methods.
- The proposed interactive face annotation is about 1.73x faster than manual annotation and 1.32x faster than segment-based methods; texture annotation reaches 87.9 mIoU with 49.3 boundary mIoU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use face-centered sampling instead of other sampling strategies (ablation) | 3D semantic segmentation on the face labeling track · Face labeling track | mIoU: Average mIoU: 48.0 (Fc.), 40.7 (Rd.), 38.2 (Po.), 44.0 (Sp.) | Helped |
| Use superpixel texture sampling instead of other sampling strategies (ablation) | 3D semantic segmentation on the pixel labeling track · Pixel labeling track | mIoU: Average mIoU: 30.6 (Rd.), 27.8 (Po.), 31.5 (Sp.) | Helped |
| Use PointVector with face-centered sampling | 3D semantic segmentation · Face labeling track | mIoU: 70.0 | Helped |
| Use PointVector with superpixel texture sampling | 3D semantic segmentation · Pixel labeling track | mIoU: 47.9% | Helped |
| Use the proposed interactive face annotation method instead of manual annotation | Interactive face annotation · Face annotation | T(s): 4183.2 -> 2992.1 | Helped |
| Use the proposed interactive face annotation method instead of segment-based annotation | Interactive face annotation · Face annotation | T(s): 3565.8 -> 2992.1 | Helped |
| Use the proposed texture annotation method instead of GrabCut | Interactive texture annotation · Texture annotation | B (%): 47.1 -> 49.3 | Helped |
| Use the proposed texture annotation method instead of SAM | Interactive texture annotation · Texture annotation | T(s): 640.6 -> 663.3 | Mixed |
| Increase the face-based balance parameter λf (ablation) | Proposed face annotation method · Face annotation | object boundary clarity | Helped |
| Increase face segment-matching thresholds ϵ(seg) and ϵ(str) (ablation) | Proposed face annotation method · Face annotation | coverage of repetitive structures | Helped |
| Add Gaussian noise to mesh input (ablation) | Proposed face annotation method · Protrusion extraction | protrusion extraction robustness | No clear effect |
| Increase the texture-based balance parameter λs (ablation) | Proposed texture annotation method · Texture annotation | local-region smoothness | Helped |
Surprising findings
- Face-centered point sampling gives the best segmentation for most methods, outperforming the proposed superpixel texture sampling on face labeling.
Limitations
- Annotation method relies on geometric precision and structural clarity; struggles with topological errors, low resolution, or poor planar over-segmentation.
- Not applicable to natural scenes or complex non-planar structures.
- Texture annotation degrades with complex textures, cluttered backgrounds, shadows, or minimal color differences.
Experimental setup
Datasets: SUM Parts
Benchmarks: SUM Parts
Metrics: IoU overall accuracy mean accuracy mIoU mean IoU boundary mean IoU average user interactions annotation time smart interaction ratio
Baselines: PointNet PointNet++ SPG SparseUNet RandLA-Net KPConv PointNext PointTransV3 PointVector RF-MRF SUM-RF PSSNet
Builds on: GrabCut SLIC Point Transformer V3 PointNext KPConv
Closest papers by meaning
- Towards Semantic Segmentation of Urban-Scale 3D Point Clouds: A Dataset, Benchmarks and Challenges (CVPR 2021 · similarity 0.89)
- Part-Aware Panoptic Segmentation (CVPR 2021 · similarity 0.89)
- PartNet: A Large-Scale Benchmark for Fine-Grained and Hierarchical Part-Level 3D Object Understanding (CVPR 2019 · similarity 0.89)
- BFANet: Revisiting 3D Semantic Segmentation with Boundary Feature Analysis (CVPR 2025 · similarity 0.87)
- PartNeXt: A Next-Generation Dataset for Fine-Grained and Hierarchical 3D Part Understanding (NeurIPS 2025 · similarity 0.87)
- LogoSP: Local-global Grouping of Superpoints for Unsupervised Semantic Segmentation of 3D Point Clouds (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.