Abstract
Non-rigid point cloud registration often fails during topological changes, such as objects separating, due to the "sticky" nature of spatial smoothness constraints. We present FB-Warp, a pipeline that leverages time symmetry to detect topological events and blend motion hypotheses. By identifying regions of stretching and compression between forward and backward warps, our method achieves state-of-the-art accuracy on dynamic scenes, reducing error by 30% at separation boundaries.
Introduction
Non-rigid registration is a foundational task for 3D reconstruction and robotics, yet it consistently struggles with "topology changes"—events where surfaces touch or separate. Standard algorithms typically rely on $k$-nearest neighbor ($k$-NN) graphs to ensure motion smoothness. However, when objects move apart, these graphs often "stick" the surfaces together, causing the motion estimate to smear across the newly formed gap.
While volumetric approaches using Signed Distance Fields can mitigate this, they are computationally expensive and do not scale well. We propose a topology-aware approach that operates directly on unorganized point clouds, ensuring efficiency and scalability for large-scale scenes.
Methodology: The FB-Warp Pipeline
The core innovation of our approach, FB-Warp, is a post-processing phase that utilizes the symmetry of forward and backward motion to resolve topological ambiguities.
- Dual Warp Hypotheses: We compute both a standard forward warp ($S \to D$) and a backward warp ($D \to S$). A "separation" in forward time is a "contact" in backward time. Because standard algorithms handle contacts effectively (as no gap exists to bridge), the inverted backward warp provides a "clean" hypothesis at separation boundaries.
- Event Detection: The algorithm identifies topological events by analyzing local "neighborhood stretching" and "compression." Significant stretching in the forward warp indicates a separation, while compression indicates a contact.
- Local Blending: We use a Radial Basis Function (RBF) kernel to smoothly blend the two hypotheses. The system trusts the inverted backward warp near separations and the standard forward warp near contacts or stable regions.
Results and Evaluation
We evaluated FB-Warp on the MPI Sintel dataset and a custom "Topological Event" dataset featuring manual annotations for contacts and separations (e.g., hand-clapping, opening drawers).
- State-of-the-Art Accuracy: FB-Warp achieved the lowest overall Endpoint Error (EPE) on the MPI Sintel dataset compared to baselines like PD-Flow.
- Significant Error Reduction: In regions undergoing "close-to-open" (separation) changes, the method reduced registration error by approximately 30% compared to topology-agnostic baselines.
- Robust Event Detection: The system achieved an F-score of 0.8 in explicitly detecting topological events, proving its utility for higher-level scene understanding.
Conclusion
FB-Warp effectively addresses the "sticky" boundary problem in non-rigid registration by exploiting time symmetry. While challenges remain with sliding contacts where local distances do not change significantly, the method provides a robust, topology-aware framework for dynamic 3D reconstruction without requiring pre-existing models or semantic segmentation.
Summary updated