Research map / Point clouds / Point cloud classification
CamPoint: Boosting Point Cloud Segmentation with Virtual Camera
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.
Proposes CamPoint, which uses virtual cameras to build per-point visibility features, a camera perspective slice distance for semantic neighbor search, and learnable camera embeddings fed into a Mamba-style SSM for global point cloud segmentation.
Problem
Point cloud segmentation struggles with selecting semantically relevant neighbors (Euclidean KNN is spatially biased) and with providing high-level, stable features for global perception at low computational cost.
Approach
Virtual cameras produce a binary visibility vector per point; the number of shared cameras between two points defines a camera perspective slice distance that is combined with Euclidean distance for KNN, and learnable per-camera embeddings are pooled over visible cameras and added to local features before a random-scan SSM.
Key results
- CamPoint reaches 83.3% S3DIS Area 5 mIoU and 77.7% ScanNetV2 val mIoU, outperforming PTv3 by 8.6 and 0.2 points respectively.
- On ShapeNetPart, CamPoint achieves 87.2 instance mIoU versus 86.9 for PCM.
- Ablation on S3DIS: adding camera perspective slice distance raises Local baseline from 72.9 to 76.2 mIoU; full model reaches 83.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add camera perspective slice distance to local aggregation (ablation) | CamPoint · S3DIS | mIoU: 72.9 -> 76.2 | Helped |
| Add camera perspective slice distance to local aggregation (ablation) | CamPoint · ScanNetV2 | mIoU: 74.8 -> 75.5 | Helped |
| Add global SSM perception after local aggregation with CPSD (ablation) | CamPoint · S3DIS | mIoU: 76.2 -> 82.1 | Helped |
| Add global SSM perception after local aggregation with CPSD (ablation) | CamPoint · ScanNetV2 | mIoU: 75.5 -> 77.0 | Helped |
| Add camera parameter embedding to global perception (ablation) | CamPoint · S3DIS | mIoU: 82.1 -> 83.3 | Helped |
| Add camera parameter embedding to global perception (ablation) | CamPoint · ScanNetV2 | mIoU: 77.0 -> 77.7 | Helped |
| Add camera perspective slice distance (ablation) | PointNext-s · S3DIS | mIoU: 63.4 -> 64.9 | Helped |
| Add camera perspective slice distance (ablation) | PointNext-s · ScanNetV2 | mIoU: 64.5 -> 65.3 | Helped |
| Add camera perspective slice distance (ablation) | DeLA · S3DIS | mIoU: 74.1 -> 77.8 | Helped |
| Add camera perspective slice distance (ablation) | DeLA · ScanNetV2 | mIoU: 75.9 -> 76.7 | Helped |
| Increase camera count from 2×32 to 2×64 (ablation) | CamPoint · S3DIS | mIoU: 81.4 -> 83.3 | Helped |
| Increase camera count from 2×32 to 2×64 (ablation) | CamPoint · ScanNetV2 | mIoU: 76.4 -> 77.7 | Helped |
Surprising findings
- Camera perspective slice distance also improves PointNeXt-s and DeLA when plugged in, suggesting the camera prior is a generalizable neighbor-search signal.
Limitations
- Relies on predefined camera placements, which may limit adaptability in complex environments.
- Does not handle segmentation in dynamic scenes.
- Gains on ShapeNetPart are modest given the strong baseline.
Experimental setup
Datasets: S3DIS ScanNetV2 ShapeNetPart ScanObjectNN ModelNet40
Benchmarks: S3DIS ScanNetV2 ShapeNetPart ScanObjectNN ModelNet40
Metrics: mIoU instance mIoU class mIoU overall accuracy mean accuracy parameters FLOPs inference latency training time
Baselines: MinkUNet PointNeXt DeLA PointHR OA-CNNs KPConvX-L Stratified Transformer OctFormer Swin3D PointNAT OneFormer PTv3
Builds on: DeLA Mamba PointNeXt Point Mamba
Closest papers by meaning
- A Linear Time and Space Local Point Cloud Geometry Encoder via Vectorized Kernel Mixture (VecKM) (ICML 2024 · similarity 0.86)
- FreePoint: Unsupervised Point Cloud Instance Segmentation (CVPR 2024 · similarity 0.86)
- Taylor Series-Inspired Local Structure Fitting Network for Few-shot Point Cloud Semantic Segmentation (AAAI 2025 · similarity 0.85)
- Geometrically-driven Aggregation for Zero-shot 3D Point Cloud Understanding (CVPR 2024 · similarity 0.85)
- X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition (CVPR 2024 · similarity 0.85)
- Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering (AAAI 2025 · similarity 0.85)
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.