UniRiT: Towards Few-Shot Non-Rigid Point Cloud Registration
Geng Li, Haozhi Cao, Mingyang Liu, Chenxi Jiang, Jianfei Yang
TL;DR
UniRiT introduces a two-step framework for few-shot non-rigid point cloud registration, effectively handling complex transformations with limited data, validated on a new real-world dataset showing significant performance improvements.
Contribution
The paper proposes UniRiT, a novel two-step registration method that decomposes non-rigid transformations, and introduces MedMatch3D, a real-world dataset for few-shot registration benchmarking.
Findings
Achieves state-of-the-art performance on MedMatch3D
Improves existing methods by 94.22%
Effectively handles complex non-rigid transformations with limited data
Abstract
Non-rigid point cloud registration is a critical challenge in 3D scene understanding, particularly in surgical navigation. Although existing methods achieve excellent performance when trained on large-scale, high-quality datasets, these datasets are prohibitively expensive to collect and annotate, e.g., organ data in authentic medical scenarios. With insufficient training samples and data noise, existing methods degrade significantly since non-rigid patterns are more flexible and complicated than rigid ones, and the distributions across samples are more distinct, leading to higher difficulty in representation learning with few data. In this work, we aim to deal with this challenging few-shot non-rigid point cloud registration problem. Based on the observation that complex non-rigid transformation patterns can be decomposed into rigid and small non-rigid transformations, we propose a…
Peer Reviews
Decision·ICLR 2025 Conference Withdrawn Submission
1. Propose a straight-forward two-step shape registration method to match the point cloud from global alignment to local matching using learned transformation. 2. Provide a dataset for 3D organ shapes
1. The authors spent much space on explaining the gaussian mixture model and tried to related the point cloud registration with GMM as in previous works. However, in the later methodology part, the introduced network structure and the loss function has no direct relation to the GMM model. The network structure shares similarity with the point cloud segmentation models, while the loss function is similar to chamfer distance. Please explicitly explain how the GMM analysis informed their network de
- UniRiT's two-step decomposition simplifies the complex problem of N-PCR, improving the feasibility of few-shot learning. - Demonstrated substantial performance gains, compared to optimization-based and learning-based N-PCR methods. - The introduction of MedMatch3D provides a valuable benchmark for future research in medical applications of N-PCR. - The paper demonstrates that UniRiT achieves competitive registration accuracy while maintaining lower computational complexity and faster runtime c
- The two-stage registration method—rigid followed by non-rigid—is a well-established practice in image and shape registration, particularly in medical applications. This limits the novelty of UniRiT’s methodological framework. - The framework assumes that the source and target point clouds have the same number of points, which may not hold in real-world applications. It is unclear why this restriction is necessary, especially since real-world point clouds often differ in size unless resampled.
The main contribution of the paper is introducing a new challenging problem for the researchers to look into this problem. By creating a new benchmark dataset, it will allow easier performance evaluation and comparison for the researchers working on this problem.
Although it is always beneficial to have new datasets, this dataset is primarily derived from a subset of an existing MedShapeNet dataset, followed by some cleaning and the application of Thin Plate Splines (TPS) to create pairs for registration. It does not add significant value to the original dataset, as the most valuable component—the real CT or MR human anatomical data—originates from MedShapeNet. I would appreciate more visuals of the source and target organs point clouds, but based on t
The paper presents a method to address non-rigid point cloud registration on point clouds of organs. This scenario is quite different and more challenging than deformable point-cloud registration on benchmark point clouds, e.g. [1], and also, it is not as widely studied. Additionally, the challenge is entailed in the fact that the intra- and inter-organ deformations are way more challenging to recover and that the medical point cloud data can be scarce. The paper also proposes and constructs
Although I believe the paper is very interesting and attempts to address very challenging and significant tasks, I would also like to point out several weaknesses. [A] The second contribution indicates that the paper presents a method that recovers the sought transformation in a rigid and non-rigid step, simplifying the learning process. However, I believe this is a well-studied concept in point-cloud and image registration, and many other methods first estimate a rough rigid alignment before t
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsRemote Sensing and LiDAR Applications · 3D Surveying and Cultural Heritage · Robotics and Sensor-Based Localization