Research map / Point clouds / Point cloud classification
PointDGMamba: Domain Generalization of Point Cloud Classification via Generalized State Space Model
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 PointDGMamba, a state-space-model (Mamba) framework for domain-generalized point cloud classification with masked sequence denoising, cross-domain feature aggregation and dual-level domain scanning, plus a new PointDG-3to1 multi-domain benchmark.
Problem
Domain generalization for point cloud classification: CNNs have limited receptive fields and ViTs have quadratic cost, while existing Mamba-based point models generalize poorly to unseen domains due to serialization noise and domain-specific scanning.
Approach
Adapts Mamba to DG point cloud classification with three components: Gumbel-Softmax learnable masking to filter noisy point tokens (MSD), aggregation of same-class features from other source domains plus a learnable global prompt (SCFA), and intra-domain plus cross-domain scanning orders (DDS).
Key results
- On PointDA-10, PointDGMamba reaches 74.85% average accuracy, 2.94% above MetaSets (71.91%).
- On PointDG-3to1, PointDGMamba reaches 80.53% average, exceeding PDG (79.21%) by 1.32%.
- Ablation: SCFA, MSD and DDS each add gains; full model improves over the no-module baseline (70.71%) to 74.85% average on PointDA-10.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add cross-domain features (CDF) to SCFA (ablation) | PointDGMamba · PointDA-10 | accuracy: 70.71 -> 72.75 | Helped |
| Add global prompt (GP) to SCFA (ablation) | PointDGMamba with CDF · PointDA-10 | accuracy: 72.75 -> 73.23 | Helped |
| Add MSD to SCFA with CDF and GP (ablation) | PointDGMamba · PointDA-10 | accuracy: 73.23 -> 74.85 | Helped |
| Add IDS and CDS together (ablation) | PointDGMamba without scanning · PointDA-10 | accuracy: 72.85 -> 74.85 | Helped |
| Replace MSD with random masking (ablation) | PointDGMamba · PointDA-10 | accuracy: 74.85 -> 69.75 | Hurt |
| Replace MSD with similarity masking (ablation) | PointDGMamba · PointDA-10 | accuracy: 74.85 -> 70.98 | Hurt |
| Replace MSD with IIM masking (ablation) | PointDGMamba · PointDA-10 | accuracy: 74.85 -> 69.07 | Hurt |
| Replace DDS with forward scanning (ablation) | PointDGMamba · PointDA-10 | accuracy: 74.85 -> 72.87 | Hurt |
| Replace DDS with backward scanning (ablation) | PointDGMamba · PointDA-10 | accuracy: 74.85 -> 73.29 | Hurt |
| Replace DDS with shuffled scanning (ablation) | PointDGMamba · PointDA-10 | accuracy: 74.85 -> 72.59 | Hurt |
| Use PointDGMamba-Base with two Mamba stages (ablation) | PointDGMamba-Tiny · PointDA-10 | accuracy: 74.85 -> 73.70 | Hurt |
| Reduce feature dimension from 192 to 128 (ablation) | PointDGMamba-Tiny · PointDA-10 | accuracy: 74.85 -> 74.21 | Hurt |
Surprising findings
- Randomly masking 5% of sequence tokens or using a similarity-based mask gives lower gains than learned MSD, and Fourier-based FDA aggregation performs worst among aggregation methods tested.
Limitations
- Training requires cropping long point cloud sequences, which may hurt feature representativeness.
- Only evaluated on classification; segmentation and large-scale scenes are left for future work.
- Baselines not designed for DG were modified by the authors to fit the DG protocol.
Experimental setup
Datasets: PointDA-10 ModelNet ShapeNet ScanNet 3D-FUTURE
Benchmarks: PointDA-10 PointDG-3to1
Metrics: classification accuracy average accuracy across target domains parameters (M) GFLOPs inference time (ms)
Baselines: PointDAN DefRec GAST PDG MetaSets PointNeXt X-3D PCT GBNet SUG PCM
Builds on: Mamba PCM Gumbel-Softmax PointMix
Closest papers by meaning
- DAPoinTr: Domain Adaptive Point Transformer for Point Cloud Completion (AAAI 2025 · similarity 0.88)
- PointMamba: A Simple State Space Model for Point Cloud Analysis (NeurIPS 2024 · similarity 0.88)
- Point-PRC: A Prompt Learning Based Regulation Framework for Generalizable Point Cloud Analysis (NeurIPS 2024 · similarity 0.87)
- Point Cloud Mamba: Point Cloud Learning via State Space Model (AAAI 2025 · similarity 0.87)
- PointRWKV: Efficient RWKV-Like Model for Hierarchical Point Cloud Learning (AAAI 2025 · similarity 0.86)
- ZigzagPointMamba: Spatial-Semantic Mamba for Point Cloud Understanding (NeurIPS 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.