ERNet: Efficient Non-Rigid Registration Network for Point Sequences

TL;DR

The key to the design is predicting a sequence of deformation graphs through a two-stage pipeline, which first estimates frame-wise coarse graph nodes for robust initialization, before refining their trajectories over time in a sliding-window fashion.

Abstract

Registering an object shape to a sequence of point clouds undergoing non-rigid deformation is a long-standing challenge. The key difficulties stem from two factors: (i) the presence of local minima due to the non-convexity of registration objectives, especially under noisy or partial inputs, which hinders accurate and robust deformation estimation, and (ii) error accumulation over long sequences, leading to tracking failures. To address these challenges, we introduce to adopt a scalable data-driven approach and propose ERNet, an efficient feed-forward model trained on large deformation datasets. It is designed to handle noisy and partial inputs while effectively leveraging temporal information for accurate and consistent sequential registration. The key to our design is predicting a sequence of deformation graphs through a two-stage pipeline, which first estimates frame-wise coarse graph nodes for robust initialization, before refining their trajectories over time in a sliding-window fashion. Extensive experiments show that our proposed approach (i) outperforms previous state-of-the-art on both the DeformingThings4D and D-FAUST datasets, and (ii) achieves more than 4x speedup compared to the previous best, offering significant efficiency improvement.

Figures (3)

• Figure 1: Overview of our proposed pipeline. (a) Given a source point cloud and input target point cloud sequence, we first encode them independently using a shared local feature encoder and splat per-point features onto triplane grids (Sec. \ref{['method:feature']}). (b) Then we initialize graph nodes based on source point cloud and perform the coarse-to-fine matching to predict the node positions and radii of the deformation graph using encoded features (Sec. \ref{['method:transformer']}). (c) With the predicted node trajectories and radii, we calculate node transformations via the Procrustes analysis and utilize the RBF-based LBS to produce dense registration (Sec. \ref{['method:skinning']}).
• Figure 2: Qualitative results on two challenging examples from the depth-sampled D-FAUST and DT4D-A datasets. The point color reflects the L2 distance from ground truth, where blue indicates less error and red indicates more.
• Figure 3: Analysis on the effect of update iteration and node count. Metrics $ATE_{3D}$, $\delta_{0.01}$ and $\delta_{0.05}$ w.r.t. transformer update iterations and node counts are presented. We update trajectories 6 times and set node count $B=256$ as a balance between performance and efficiency.