Research map / Point clouds / Point cloud classification
KPConvX: Modernizing Kernel Point Convolution with Kernel Attention
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 KPConvD (depthwise kernel point convolution) and KPConvX (adds geometric kernel attention with group-wise modulations), combined with a modern inverted-bottleneck architecture and training recipe, achieving state-of-the-art point cloud segmentation and classification.
Problem
Kernel Point Convolution (KPConv) has been surpassed by recent MLP-based and transformer point cloud networks; the goal is to modernize KPConv for efficient and accurate 3D point cloud understanding.
Approach
Replace KPConv's full weight matrices with depthwise kernel weights combined with the Hadamard product, and modulate them with geometric kernel attention generated from the central point's feature (one modulation vector per kernel point, grouped), using nearest-kernel assignment and a two-shell kernel layout.
Key results
- KPConvX-L reaches 73.5 mIoU on S3DIS Area 5, +1.2 over the prior best (StratTrans 72.0 / PointVector 72.3 per Table 1).
- KPConvX-L achieves 76.3 mIoU on ScanNetv2 validation, outperforming OctFormer (75.7) by +0.6.
- In controlled ablation on S3DIS, KPConvX-L averages 72.4 mIoU vs 68.5 for the reimplemented KPConv (+3.9), running 35% faster with half the GPU memory.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KPConvX-L instead of StratTrans | S3DIS Area 5 semantic segmentation · S3DIS Area 5 | mIoU: 72.0 -> 73.5 | Helped |
| Use KPConvX instead of OctFormer | ScanNetv2 semantic segmentation · ScanNetv2 validation | mIoU: 75.7 -> 76.3 | Helped |
| Use KPNeXt-L instead of PointVector | ScanObjectNN PB_T50_RS classification · ScanObjectNN PB_T50_RS | OA: 87.8 ± 0.4 -> 88.9 ± 0.3 | Helped |
| Add depthwise convolution (ablation) | KPConv · S3DIS Area 5 | mIoU: 68.5 ± 0.2 -> 68.0 ± 0.3 | Hurt |
| Reduce the radius and use a fixed number of neighbors (ablation) | KPConvD · S3DIS Area 5 | mIoU: 68.0 ± 0.3 -> 68.2 ± 0.3 | Helped |
| Add a second kernel shell, using [1, 14, 28] kernel points (ablation) | KPConvD · S3DIS Area 5 | mIoU: 68.2 ± 0.3 -> 69.0 ± 0.6 | Helped |
| Use the new large architecture (ablation) | KPConvD · S3DIS Area 5 | mIoU: 69.0 ± 0.6 -> 70.9 ± 0.7 | Helped |
| Use inverted bottleneck blocks (ablation) | KPConvD · S3DIS Area 5 | mIoU: 70.9 ± 0.7 -> 72.2 ± 0.4 | Helped |
| Use nearest-kernel influence (ablation) | KPConvD · S3DIS Area 5 | mIoU: 72.2 ± 0.4 -> 72.3 ± 0.4 | No clear effect |
| Share kernel points across blocks (ablation) | KPConvD-L · S3DIS Area 5 | mIoU: 72.3 ± 0.4 -> 72.2 ± 0.7 | No clear effect |
| Add kernel attention (ablation) | KPConvD-L · S3DIS Area 5 | mIoU: 72.2 ± 0.7 -> 72.4 ± 0.9 | No clear effect |
| Use the large rather than small architecture (ablation) | KPConvX · S3DIS Area 5 | mIoU: 71.1 ± 0.5 -> 72.4 ± 0.9 | Helped |
Surprising findings
- The depthwise KPConvD-L nearly matches KPConvX-L (72.2 vs 72.4 mIoU) while being about 35% faster, suggesting the attention adds limited gain on S3DIS.
- Using a channel-width ratio of sqrt(2) between layers instead of 2 did not hurt performance while reducing size and memory.
Limitations
- Kernel attention is only applied to a geometric operator; the authors note open questions about combining topological and geometric features.
- Results rely on test-time voting and a modern training recipe, so gains partly come from training strategy rather than the operator alone.
- ScanObjectNN results are reported without pre-training and compared only against non-pretrained methods.
Experimental setup
Datasets: S3DIS ScanNetv2 ScanObjectNN
Benchmarks: S3DIS Area 5 ScanNetv2 ScanObjectNN PB_T50_RS
Metrics: mIoU mAcc OA
Baselines: KPConv PTv1 PTv2 StratTrans PointVector PointNeXt OctFormer SPoTr PointMLP PointStack
Builds on: KPConv Image involution PointNeXt Deformable KPConv Point Transformer V2
Closest papers by meaning
- FPConv: Learning Local Flattening for Point Convolution (CVPR 2020 · similarity 0.87)
- RepKPU: Point Cloud Upsampling with Kernel Point Representation and Deformation (CVPR 2024 · similarity 0.84)
- Improved MLP Point Cloud Processing with High-Dimensional Positional Encoding (AAAI 2024 · similarity 0.83)
- DeepLA-Net: Very Deep Local Aggregation Networks for Point Cloud Analysis (CVPR 2025 · similarity 0.83)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.83)
- CAM Back Again: Large Kernel CNNs from a Weakly Supervised Object Localization Perspective (CVPR 2024 · similarity 0.83)
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.