TL;DR
DAPointRWKV is the first framework to adapt the efficient Receptance Weighted Key Value (RWKV) architecture for domain-adaptive 3D point cloud completion. By combining geometry-aware recurrent modeling with a novel dual-stabilization strategy, it achieves state-of-the-art accuracy across synthetic and real-world datasets while maintaining linear computational complexity.
What Problem Does This Paper Address?
Point cloud completion (PCC) is the task of predicting a complete 3D shape from a partial, often noisy, scan. While supervised methods work well, they rely on paired "partial-complete" data that is nearly impossible to collect in the real world. This necessitates Unsupervised Domain Adaptation (UDA), where a model learns from labeled synthetic data and adapts to unlabeled real-world scans.
Existing UDA methods face a "trilemma" of challenges:
- Computational Cost: Transformers offer great performance but suffer from quadratic complexity ($O(L^2)$), making them slow for dense point clouds.
- Geometric Bias: Efficient sequence models like Mamba often struggle to capture the local spatial relationships of 3D points once they are flattened into 1D sequences.
- Adaptation Instability: Without labels in the target domain, models often produce "hallucinated" or degenerate shapes that don't match the original partial observation.
Methods
The authors propose DAPointRWKV, which introduces three tailored components to bridge the gap between efficient sequence modeling and 3D geometry:
- Geometry-Aware RWKV (GA-RWKV) Backbone: Standard RWKV blocks are modified with "geometry gates." These gates inject local structural descriptors (based on relative coordinates and distances) directly into the recurrent mixing process, ensuring the model doesn't lose spatial context when processing serialized tokens.
- RWKV-Oriented Progressive Decoder (RWKV-PD): To bridge the gap between 1D features and 3D shapes, this decoder uses a coarse-to-fine approach. It generates an initial "seed" cloud and then progressively refines it using local token-guided features retrieved from the backbone.
- Dual Structural Regularization (DSR): This is the core adaptation engine. It uses two constraints to stabilize learning on unlabeled data:
- Observation-Anchor Constraint (OAC): Ensures the predicted complete shape still contains the original partial points.
- Hypothesis-Stabilizing Constraint (HSC): A self-consistency check where the model's own prediction is re-cropped and fed back in; the model must produce a consistent full shape regardless of which partial view it sees.
Datasets
The model was trained on the CRN (synthetic) source dataset and evaluated on several challenging target domains:
- Synthetic Targets: 3D-FUTURE (furniture) and ModelNet (general objects).
- Real-World Targets: ScanNet (indoor scenes), MatterPort3D (indoor environments), and KITTI (outdoor driving scenes).
Key Insights and Ideas
- Linear Complexity is King: As shown in the paper's trade-off analysis, DAPointRWKV scales linearly ($O(L)$) with the number of points, making it significantly more efficient than Transformer-based SOTA like DAPoinTr as the resolution increases.
- Inductive Bias Matters: Simply applying a "vision" model to 3D isn't enough. The inclusion of the Geometry Gate in the backbone was crucial for handling irregular point distributions.
- Stabilizing the "Hallucination": The Dual Structural Regularization (DSR) solves a major headache in UDA. By forcing the model to be consistent across different "views" of its own predictions (HSC), it prevents the model from converging to degenerate solutions during unsupervised training.
Key Results
DAPointRWKV consistently outperformed previous SOTA methods (DAPoinTr and DAPointMamba) across all benchmarks:
- Accuracy: On the 3D-FUTURE benchmark, it reduced the average Chamfer Distance (CD) to 16.02, a significant improvement over DAPointMamba's 20.40.
- Efficiency: It uses roughly half the parameters (22.48M vs 48.23M) and is nearly 5x faster in inference time (4.98ms vs 24.00ms) than Transformer-based competitors.
- Real-World Robustness: It achieved the best results on the KITTI "Car" category, proving its effectiveness on sparse, noisy outdoor data.
Limitations and Future Work
The authors note that while the Observation-Anchor Constraint (OAC) is vital, if used alone, it can bias the model toward being too "conservative"—essentially just outputting the partial input without filling in the gaps. The balance between OAC and HSC is delicate.
Future work could involve exploring even more complex "re-observation" strategies for the DSR module or extending the architecture to handle full-scene completion rather than single-object completion.
Summary updated