ERNet: Efficient Non-Rigid Registration Network for Point Sequences

Guangzhao He, Yuxi Xiao, Zhen Xu, Xiaowei Zhou, Sida Peng
Zhejiang University
ICCV 2025
Paper arXiv Code
ERNet Teaser

ERNet efficiently and accurately predicts feed-forward registrations given a source object and a sequence of sparse or partial point clouds.

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 propose 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 DeformingThings4D and D-FAUST datasets, and (ii) achieves more than 4x speedup compared to the previous best, offering significant efficiency improvement.

Method

ERNet Method

(a) Given a source point cloud and input target point cloud sequence, ERNet first encodes them independently using a shared local feature encoder and splat per-point features onto Tri-plane grids.

(b) Deformation graph nodes are initialized based on the source point cloud and coarse-to-fine matching is performed to predict node positions and radii using encoded features.

(c) With the predicted node trajectories and radii, node transformations are computed via Procrustes Analysis and dense registration is produced using RBF-based LBS.

Registration Results

ERNet Method

Qualitative comparison results on two challenging examples from partial 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.

Registration Results

ERNet Method

Additional qualitative results from D-FAUST and DT4D-A datasets. The error heatmap ranges from 0 to 0.4, with 0 indicating perfect registration and 0.4 indicating the worst.

BibTeX

@article{he2025ernet,
  author    = {He, Guangzhao and Xiao, Yuxi and Xu, Zhen and Zhou, Xiaowei and Peng, Sida},
  title     = {ERNet: Efficient Non-Rigid Registration Network for Point Sequences},
  journal   = {ICCV},
  year      = {2025},
}