Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MATTER: Multiscale Attention for Registration Error Regression

Shipeng Liu, Ziliang Xiong, Khac-Hoang Ngo, Per-Erik Forssén

TL;DR

This paper tackles the problem of estimating registration quality for point-cloud pairs by reframing misalignment detection as a regression task. It introduces MATTER, a multiscale cross-attention framework that fuses differential entropy, Sinkhorn divergence, and reliability features across multiple radii to predict the alignment error $E_{ ext{align}}$. Across three challenging datasets, MATTER surpasses prior classification-based methods in RMSE, MAE, and $R^2$, and remains robust under poor registrations and low overlap. When applied to a mapping pipeline, MATTER improves downstream map quality at a fixed re-registration budget, demonstrating practical impact for SLAM and 3D tracking tasks.

Abstract

Point cloud registration (PCR) is crucial for many downstream tasks, such as simultaneous localization and mapping (SLAM) and object tracking. This makes detecting and quantifying registration misalignment, i.e., PCR quality validation, an important task. All existing methods treat validation as a classification task, aiming to assign the PCR quality to a few classes. In this work, we instead use regression for PCR validation, allowing for a more fine-grained quantification of the registration quality. We also extend previously used misalignment-related features by using multiscale extraction and attention-based aggregation. This leads to accurate and robust registration error estimation on diverse datasets, especially for point clouds with heterogeneous spatial densities. Furthermore, when used to guide a mapping downstream task, our method significantly improves the mapping quality for a given amount of re-registered frames, compared to the state-of-the-art classification-based method.

MATTER: Multiscale Attention for Registration Error Regression

TL;DR

This paper tackles the problem of estimating registration quality for point-cloud pairs by reframing misalignment detection as a regression task. It introduces MATTER, a multiscale cross-attention framework that fuses differential entropy, Sinkhorn divergence, and reliability features across multiple radii to predict the alignment error . Across three challenging datasets, MATTER surpasses prior classification-based methods in RMSE, MAE, and , and remains robust under poor registrations and low overlap. When applied to a mapping pipeline, MATTER improves downstream map quality at a fixed re-registration budget, demonstrating practical impact for SLAM and 3D tracking tasks.

Abstract

Point cloud registration (PCR) is crucial for many downstream tasks, such as simultaneous localization and mapping (SLAM) and object tracking. This makes detecting and quantifying registration misalignment, i.e., PCR quality validation, an important task. All existing methods treat validation as a classification task, aiming to assign the PCR quality to a few classes. In this work, we instead use regression for PCR validation, allowing for a more fine-grained quantification of the registration quality. We also extend previously used misalignment-related features by using multiscale extraction and attention-based aggregation. This leads to accurate and robust registration error estimation on diverse datasets, especially for point clouds with heterogeneous spatial densities. Furthermore, when used to guide a mapping downstream task, our method significantly improves the mapping quality for a given amount of re-registered frames, compared to the state-of-the-art classification-based method.

Paper Structure

This paper contains 16 sections, 9 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Task Formulation
  4. Model Architecture
  5. Feature Extraction
  6. Multiscale Cross-Attention
  7. Evaluation
  8. Datasets for Evaluation
  9. Evaluation Results
  10. Downstream Task
  11. Conclusion
  12. Training Settings
  13. Related Works
  14. Feature Extraction
  15. Ablation on Temperature
  16. ...and 1 more sections

Figures (5)

  • Figure 1: Overview of the proposed multiscale attention mechanism for geometric features in registered point clouds.
  • Figure 2: Visualization of the registered point clouds. Left: reference (blue) vs. estimated-aligned source (red). Right: scale preference on anchor points.
  • Figure 3: Visualization of re-registered point cloud sequences. Left: FACT, $E_{\text{align}} = 21.46$ m; Right: MATTER, $E_{\text{align}} = 18.35$ m.
  • Figure 4: The final-frame alignment error vs. re-registration rate when using MATTER/FACT for misalignment detection.
  • Figure 5: Larger visualizations from the main paper. Top: registered point clouds and scale preference. Bottom: re-registered point cloud sequences (Left: FACT, $E_{\text{align}} = 21.46$ m; Right: MATTER, $E_{\text{align}} = 18.35$ m).