Research map / Point clouds / Point cloud classification
Point Clouds Meets Physics: Dynamic Acoustic Field Fitting Network for Point Cloud Understanding
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 DAFNet, a pretraining-free point cloud network whose DAF-Conv models local neighborhoods as multi-frequency acoustic fields split into near- and far-field weighting, reaching 96.7% OA on ScanObjectNN with 2.47M parameters.
Problem
Local structure representation in point clouds remains a bottleneck; existing pre-training and cross-modal methods do not fundamentally address it.
Approach
Treats local point aggregation as acoustic energy field propagation: neighbors are weighted by an exponential decay in the near field and a reciprocal law in the far field, and dynamic multi-frequency weights (generated from geometric priors) are combined by attention, plus a GSA layer (EdgeConv + multi-head attention) for global context.
Key results
- DAFNet reaches 96.7% overall accuracy on ScanObjectNN (PB_T50_RS) with voting, with 2.47M parameters and 1.1G FLOPs.
- On ModelNet40 DAFNet gets 94.1% OA, slightly below PointGPT's 94.9%.
- On S3DIS 2-way 1-shot few-shot segmentation DAFNet reports 69.20% mIoU, +2.79 over Seg-PN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace GPSFormer with DAFNet (with voting) | point cloud classification · ScanObjectNN PB_T50_RS | OA: 95.4 -> 96.7 | Helped |
| replace PointGPT with DAFNet (with voting) | point cloud classification · ModelNet40 | OA: 94.9 -> 94.1 | Hurt |
| replace SPoTr with DAFNet | part segmentation · ShapeNetPart | class mIoU: 85.4 -> 85.2 | Hurt |
| replace Seg-PN with DAFNet | few-shot point cloud segmentation · S3DIS 2-way 1-shot | mIoU (Avg): 66.41 -> 69.20 | Helped |
| replace Seg-PN with DAFNet | few-shot point cloud segmentation · ScanNet 2-way 1-shot | mIoU (Avg): 63.74 -> 69.37 | Helped |
| add DAF-Conv to EdgeConv + MHA (ablation) | DAFNet · ScanObjectNN PB_T50_RS | OA: 91.3 -> 96.7 | Helped |
| add MHA to EdgeConv + DAF-Conv (ablation) | DAFNet · ScanObjectNN PB_T50_RS | OA: 94.8 -> 96.7 | Helped |
| add EdgeConv to MHA + DAF-Conv (ablation) | DAFNet · ScanObjectNN PB_T50_RS | OA: 93.5 -> 96.7 | Helped |
| add neighbor coordinates pi to geometric prior [pj] (ablation) | DAFNet · ScanObjectNN PB_T50_RS | OA: 93.6 -> 94.7 | Helped |
| add relative position pj − pi to prior [pj, pi] (ablation) | DAFNet · ScanObjectNN PB_T50_RS | OA: 94.7 -> 96.2 | Helped |
| add Euclidean distance ||pj − pi|| to the geometric prior (ablation) | DAFNet · ScanObjectNN PB_T50_RS | OA: 96.2 -> 96.7 | Helped |
| increase frequency component count H from 6 to 10 (ablation) | DAFNet · ScanObjectNN PB_T50_RS | OA: 96.7 -> 95.1 | Hurt |
Surprising findings
- The pretraining-free DAFNet outperforms pre-training methods such as ReCon++ (95.25%) on ScanObjectNN.
- On ModelNet40 DAFNet trails several recent methods, which the authors attribute to saturation on synthetic data.
Limitations
- Performance on ModelNet40 is slightly below some pre-training and recent methods.
- Hyperparameters (H, d_threshold, K) are tuned on ScanObjectNN with sensitivity sweeps.
Experimental setup
Datasets: ScanObjectNN ModelNet40 ShapeNetPart S3DIS ScanNet
Benchmarks: ScanObjectNN ModelNet40 ShapeNetPart S3DIS ScanNet
Metrics: overall accuracy mean class accuracy class mIoU instance mIoU few-shot mIoU parameters FLOPs
Baselines: PointNet PointNet++ DGCNN Point-MAE ReCon++ GPSFormer Mamba3D PointMamba Seg-PN PointGPT
Builds on: PointNet++ DGCNN (EdgeConv) PAConv dynamic convolution SegPN
Closest papers by meaning
- Learning Neural Acoustic Fields (NeurIPS 2022 · similarity 0.86)
- 3DMambaIPF: A State Space Model for Iterative Point Cloud Filtering via Differentiable Rendering (AAAI 2025 · similarity 0.86)
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.85)
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 2024 · similarity 0.85)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.84)
- A Linear Time and Space Local Point Cloud Geometry Encoder via Vectorized Kernel Mixture (VecKM) (ICML 2024 · similarity 0.84)
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.