Research map / Point clouds / Registration and matching
Dual Focus-Attention Transformer for Robust Point Cloud Registration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DFAT, a coarse-to-fine point cloud registration transformer that restricts feature interaction to keypoints near superpoints and to points within the same superpoint, plus a dual-space consistency matching filter, built on GeoTransformer.
Problem
Dense feature interactions in coarse-to-fine point cloud registration (e.g., GeoTransformer) aggregate redundant or erroneous information from irrelevant points, diluting discriminability and causing wrong superpoint and point correspondences, especially under repetitive structures and low overlap.
Approach
Replace dense attention with focus-attention: at the superpoint level, interact only with top-k sparse keypoints selected from superpoint neighborhoods using feature similarity and saliency; at the point level, apply linear attention only among points belonging to the same superpoint; and filter superpoint matches by checking nearest neighbors in geometric space against feature-space matching scor
Key results
- On 3DMatch, DFAT reaches 94.9% RR and on 3DLoMatch 76.8% RR, outperforming GeoTransformer (91.5% / 74.0%) under the same LGR estimator.
- With PEAL, DFAT+PEAL reaches 95.8% on 3DMatch and 80.5% on 3DLoMatch.
- On KITTI, DFAT achieves 4.4 cm RTE, 0.19 degree RRE, and 99.8% RR, the best on all metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dual-level focus-attention and dual-space matching | GeoTransformer · 3DMatch | RR (%): 91.5 -> 94.9 | Helped |
| add dual-level focus-attention and dual-space matching | GeoTransformer · 3DLoMatch | RR (%): 74.0 -> 76.8 | Helped |
| add DFAT while using PEAL | GeoTrans+PEAL · 3DMatch | RR (%): 94.3 -> 95.8 | Helped |
| add DFAT while using PEAL | GeoTrans+PEAL · 3DLoMatch | RR (%): 78.8 -> 80.5 | Helped |
| add dual-level focus-attention and dual-space matching | GeoTransformer · KITTI | RTE (cm): 6.8 -> 4.4 | Helped |
| add dual-level focus-attention and dual-space matching | GeoTransformer · KITTI | RRE (°): 0.24 -> 0.19 | Helped |
| add dual-level focus-attention and dual-space matching | GeoTransformer · KITTI | RR (%): 99.8 -> 99.8 | No clear effect |
| add dual-level focus-attention and dual-space matching | GeoTransformer · Augmented ICL-NUIM | ATE: Living1: 17.54 -> 15.01; Living2: 13.51 -> 18.64; Office1: 13.85 -> 13.19; Office2: 9.78 -> 8.93; Mean: 14.12 -> 13.95 | Mixed |
| add superpoint focus-attention (ablation) | DFAT · 3DMatch | RR (%): 93.6 -> 94.9 | Helped |
| add dual-space matching (ablation) | DFAT · 3DMatch | RR (%): 93.1 -> 94.9 | Helped |
| add point focus-attention and dual-space matching to the superpoint-focus-only variant (ablation) | DFAT · 3DMatch | RR (%): 92.9 -> 94.9 | Helped |
| use 3 neighboring points instead of 1 in superpoint focus-attention (ablation) | DFAT · 3DMatch | RR (%): 93.0 -> 94.9 | Helped |
Surprising findings
- Ablation shows removing any single DFAT module causes a significant drop, and k=3 neighbors is best while both k=1 and k=6 reduce performance.
Limitations
- Evaluated only on rigid registration; non-rigid extension left to future work.
- Performance is sensitive to top-k neighbor count (k=3 best), suggesting hyperparameter dependence.
- Relies on GeoTransformer backbone and training losses, so gains are incremental over that base.
Experimental setup
Datasets: 3DMatch KITTI
Benchmarks: 3DMatch 3DLoMatch KITTI Augmented ICL-NUIM
Metrics: registration recall RTE RRE ATE
Baselines: GeoTransformer Predator CoFiNet RoITr BUFFER RoReg HRegNet PEAL FGR RANSAC DGR PointDSC
Builds on: GeoTransformer KPConv-FPN Predator PEAL Overlap-aware circle loss
Closest papers by meaning
- Dynamic Cues-Assisted Transformer for Robust Point Cloud Registration (CVPR 2024 · similarity 0.90)
- A Consistency-Aware Spot-Guided Transformer for Versatile and Hierarchical Point Cloud Registration (NeurIPS 2024 · similarity 0.88)
- 3D Focusing-and-Matching Network for Multi-Instance Point Cloud Registration (NeurIPS 2024 · similarity 0.88)
- FreeReg: Image-to-Point Cloud Registration Leveraging Pretrained Diffusion Models and Monocular Depth Estimators (ICLR 2024 · similarity 0.86)
- Mining and Transferring Feature-Geometry Coherence for Unsupervised Point Cloud Registration (NeurIPS 2024 · similarity 0.86)
- Learning Instance-Aware Correspondences for Robust Multi-Instance Point Cloud Registration in Cluttered Scenes (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.