Research map / Point clouds / Point cloud classification
DAPoinTr: Domain Adaptive Point Transformer for Point Cloud Completion
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 DAPoinTr, an unsupervised domain adaptation framework for point cloud completion built on PoinTr that aligns domain and point/query tokens in the Transformer encoder/decoder and uses voted prediction consistency for pseudo-labels.
Problem
Point Transformers for point cloud completion degrade when transferred from a synthetic source domain (CRN) to real or other target domains, and prior UDA completion methods rely on CNN backbones that capture only local features.
Approach
Adversarially align sequence-wise tokens inside the PoinTr encoder and decoder (domain proxy/query for global alignment, point proxy/dynamic query for token-level alignment), and treat the decoder layers as mixture-of-experts whose voted predictions give consistency constraints and pseudo-labels.
Key results
- On 3D-FUTURE, DAPoinTr reaches average CD 22.35 (x10^-4) versus 28.99 for the second-best OptDE, per the ablation table's own baseline numbers.
- On ModelNet, average CD 13.79 versus 15.94 for OptDE.
- Ablation on 3D-FUTURE: replacing PoinTr's FoldingNet with the SPD decoder lowers average CD from 33.58 to 25.74, and adding PTFA, DQFA, and VPC progressively reaches 22.35.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DAPoinTr instead of OptDE | Cross-domain point cloud completion · 3D-FUTURE | CD: 28.99 -> 22.35 | Helped |
| Use DAPoinTr instead of OptDE | Cross-domain point cloud completion · ModelNet | CD: 15.94 -> 13.79 | Helped |
| Add PTFA (ablation) | PoinTr with SPD decoder · 3D-FUTURE | CD: 25.74 -> 24.55 | Helped |
| Add DQFA (ablation) | PoinTr with SPD decoder · 3D-FUTURE | CD: 25.74 -> 23.91 | Helped |
| Add PTFA and DQFA (ablation) | PoinTr with SPD decoder · 3D-FUTURE | CD: 25.74 -> 23.26 | Helped |
| Add VPC to PTFA and DQFA (ablation) | PoinTr with SPD decoder · 3D-FUTURE | CD: 23.26 -> 22.35 | Helped |
| Add PTFA, DQFA, and VPC (ablation) | PoinTr with FoldingNet decoder · 3D-FUTURE | CD: 33.58 -> 28.47 | Helped |
| Replace FoldingNet with SPD decoder (ablation) | PoinTr · 3D-FUTURE | CD: 33.58 -> 25.74 | Helped |
| Add Domain Proxy Alignment (ablation) | DAPoinTr without Domain Proxy Alignment · 3D-FUTURE Chair | CD: 18.16 -> 18.11 | Helped |
| Add Point Proxy Alignment (ablation) | DAPoinTr without Point Proxy Alignment · 3D-FUTURE Chair | CD: 18.32 -> 18.11 | Helped |
| Add Domain Query Alignment (ablation) | DAPoinTr without Domain Query Alignment · 3D-FUTURE Chair | CD: 18.28 -> 18.11 | Helped |
| Add Dynamic Query Alignment (ablation) | DAPoinTr without Dynamic Query Alignment · 3D-FUTURE Chair | CD: 18.39 -> 18.11 | Helped |
Surprising findings
- Direct adversarial alignment on the DGCNN backbone of PoinTr gives only limited gains when transferring from CRN to ModelNet, motivating sequence-level alignment.
- Swapping FoldingNet for the SnowFlakeNet SPD decoder in the PoinTr baseline alone reduces average CD substantially (33.58 to 25.74), before any domain adaptation modules are added.
Limitations
- Still faces considerable domain gaps on objects with complex structures, with failure cases shown.
- Evaluated only on a limited set of categories per benchmark (shared categories with CRN).
Experimental setup
Datasets: CRN ScanNet MatterPort3D KITTI 3D-FUTURE ModelNet40
Benchmarks: 3D-FUTURE ModelNet ScanNet MatterPort3D KITTI
Metrics: Chamfer Distance Unidirectional Chamfer Distance Unidirectional Hausdorff Distance
Baselines: Pcl2Pcl ShapeInversion Cycle4Completion OptDE ACL-SPC DA-DGCNN
Builds on: PoinTr SnowFlakeNet DGCNN adversarial domain adaptation
Closest papers by meaning
- DAPointMamba: Domain Adaptive Point Mamba for Point Cloud Completion (AAAI 2026 · similarity 0.91)
- PointDGMamba: Domain Generalization of Point Cloud Classification via Generalized State Space Model (AAAI 2025 · similarity 0.88)
- Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · similarity 0.87)
- PointMamba: A Simple State Space Model for Point Cloud Analysis (NeurIPS 2024 · similarity 0.86)
- PointRWKV: Efficient RWKV-Like Model for Hierarchical Point Cloud Learning (AAAI 2025 · similarity 0.86)
- APML: Adaptive Probabilistic Matching Loss for Robust 3D Point Cloud Reconstruction (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.