Research map / Point clouds / Point cloud classification
Meta Architecture for Point Cloud Analysis
Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, Rongrong Ji
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use explicit position embedding instead of implicit position embedding (ablation) | Plain block with max pooling · S3DIS Area 5 | mIoU (%): 68.0±0.3 -> 69.0±0.3 | Improved |
| Use explicit position embedding instead of position pooling (ablation) | Plain block with max pooling · S3DIS Area 5 | mIoU (%): 58.4±0.6 -> 69.0±0.3 | Improved |
| Use max pooling rather than position pooling for aggregation with explicit embedding (ablation) | Plain-EPE block · S3DIS Area 5 | mIoU (%): 65.4±0.1 -> 69.0±0.3 | Improved |
| Replace vector self-attention with max pooling (ablation) | Point Transformer · S3DIS Area 5 | mIoU (%): 70.5±0.3 -> 70.3±0.2 | No clear effect |
| Add a two-layer point-update MLP to the N1P1 configuration (ablation) | Plain-EPE block · S3DIS Area 5 | mIoU (%): 69.0±0.3 -> 69.5±0.3 | Improved |
| Add an inverted bottleneck to N1P2 (ablation) | Plain-EPE block · S3DIS Area 5 | mIoU (%): 69.5±0.3 -> 69.7±0.3 | No clear effect |
| Use N1P2 rather than N0P2, adding a one-layer neighbor-update MLP (ablation) | Plain-EPE block · S3DIS Area 5 | mIoU (%): 67.8±0.4 -> 69.5±0.3 | Improved |
| Use N1P2 rather than N2P0, adding a two-layer point-update MLP and reducing neighbor-update depth (ablation) | Plain-EPE block · S3DIS Area 5 | mIoU (%): 68.7±0.3 -> 69.5±0.3 | Improved |
| Scale PointMetaBase-XXL instead of PointNeXt-XL | Point cloud segmentation models · S3DIS 6-Fold | mIoU (%): 74.9 -> 77.0 | Improved |
| Use PointMetaBase-XXL instead of PointNeXt-XL | Point cloud segmentation models · S3DIS Area-5 | mIoU (%): 70.5±0.3 -> 71.3±0.7 | Improved |
| Use PointMetaBase-S instead of PointNeXt-S | Point cloud classification models · ScanObjectNN PB_T50_RS | OA (%): 87.7±0.4 -> 87.9±0.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PointNeXt-XL — PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies (NeurIPS 2022) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual MLP Framework (ICLR 2022 · similarity 0.91)
- An Explicit Frame Construction for Normalizing 3D Point Clouds (ICML 2024 · similarity 0.86)
- Point-Cache: Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis (CVPR 2025 · similarity 0.86)
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 2024 · similarity 0.85)
- Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.85)
- CurveCloudNet: Processing Point Clouds with 1D Structure (CVPR 2024 · 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.
