Skip to content

IGASA: Integrated Geometry-Aware and Skip-Attention Modules for Enhanced Point Cloud Registration

Conference: CVPR 2026 arXiv: 2603.12719 Code: https://github.com/DongXu-Zhang/IGASA Area: 3D Vision / Point Cloud Registration / Autonomous Driving Keywords: [point cloud registration, hierarchical pyramid architecture, skip-attention, geometry-aware iterative refinement, coarse-to-fine]

TL;DR

IGASA is a point cloud registration framework that combines a Hierarchical Pyramid Architecture (HPA), Hierarchical Cross-Layer Attention (HCLA) with skip-attention fusion, and Iterative Geometry-Aware Refinement (IGAR) with dynamic consistency weighting. It achieves 94.6% Registration Recall on 3DMatch (SOTA), 100% RR on KITTI, with a total inference time of only 2.763s.

Background & Motivation

Point cloud registration (PCR) is a fundamental task for autonomous driving, robotic navigation, and environment modeling. Traditional ICP and its variants rely on nearest-neighbor distance minimization, making them sensitive to initialization and prone to local minima. Deep learning methods—particularly Transformer-based approaches such as GeoTransformer, RoITr, and SIRA-PCR—have made notable progress in capturing long-range dependencies and global context. However, a semantic gap persists: as networks deepen to acquire high-level semantics, fine-grained geometric details are diluted through downsampling. Naive skip-connection fusion strategies (concatenation or summation) fail to effectively bridge the discrepancy between low-level geometric cues and high-level semantic embeddings. Furthermore, the refinement stage of coarse-to-fine frameworks typically relies on RANSAC or hard-threshold outlier rejection, which is computationally expensive and may discard valid correspondences.

Core Problem

How to effectively bridge the semantic gap in multi-scale feature extraction—preserving fine-grained geometry while capturing global semantics—and suppress outliers more robustly during the fine registration stage?

Method

Overall Architecture

The pipeline consists of three stages: (a) the HPA module constructs a three-level feature pyramid (\(F_\text{ordinary}\), \(F_\text{minor}\), \(F_\text{primary}\)) for source point cloud \(P\) and target point cloud \(Q\), progressively enlarging the receptive field; (b) the HCLA module uses SGIRA to fuse global semantics with local geometry and SAIGA to refine features for coarse matching; (c) the IGAR module iteratively refines coarse matches by dynamically updating correspondence weights to suppress outliers, yielding the final transformation \(\{R, t\}\).

Key Designs

  1. Hierarchical Pyramid Architecture (HPA): A three-level KPConv-based encoder. The ordinary layer uses base voxel size \(dl_0\) (influence radius \(2.5 \cdot dl_0\)) to capture fine-grained local geometry; the minor layer uses voxel size \(2 \cdot dl_0\) for semi-global structure; the primary layer uses voxel size \(4 \cdot dl_0\) (convolution radius \(10 \cdot dl_0\)) for global semantics. Feature dimensions are 64, 128, and 256, respectively, with progressively increasing channel depth.
  2. Hierarchical Cross-Layer Attention (HCLA): Composed of two sub-modules—(i) SGIRA (Skip-Guided Inter-Resolution Attention): leverages primary-layer global semantic features to guide minor-layer feature fusion via a gated fusion mechanism (dual-branch convolution + adaptive gating weights + residual correction), then generates enhanced features \(F_\text{minor}^{+}\) using attention combining semantic similarity, geometric distance compensation, and skip residuals; (ii) SAIGA (Skip-Augmented Intrinsic Geometry Attention): performs self-attention on \(F_\text{minor}^{+}\), integrating geometric distance weights (controlled by learnable \(\alpha\)) and skip attention scores to sharpen the discriminability of local spatial features, yielding \(F_\text{minor}^{++}\). The two modules operate in series: SGIRA acts as a "semantic filter" and SAIGA as a "geometric sharpener."
  3. Iterative Geometry-Aware Refinement (IGAR): Introduces dynamic geometric consistency weighting in the fine registration stage, replacing hard RANSAC rejection. At each iteration \(k\), correspondence weights are recomputed based on the current transformation \((R^{(k)}, t^{(k)})\) as \(w_{ij} = \exp(-\text{residual}^2/\sigma^2) \times \mathbb{1}[\text{residual} < \tau]\), and registration error is minimized in a weighted manner. Optimal \(R^*\) and \(t^*\) are solved via weighted centroid computation and SVD of the weighted cross-covariance matrix. With \(N=5\) iterations by default, this achieves soft suppression of outliers.

Loss & Training

  • Total loss: \(L_\text{total} = L_\text{mat} + L_\text{key} + L_\text{den}\)
  • \(L_\text{mat} = \lambda_p \cdot L_p + \lambda_c \cdot L_c\): multi-level matching probability loss + weighted cross-entropy loss (distinguishing true matches from outliers)
  • \(L_\text{key} = \lambda_f \cdot L_f + \lambda_k \cdot L_k + \lambda_i \cdot L_i\): InfoNCE descriptor contrastive loss + keypoint displacement \(L_2\) loss + confidence binary cross-entropy
  • \(L_\text{den} = \lambda_t \cdot L_t + \lambda_r \cdot L_r\): translation \(L_2\) loss + rotation Frobenius norm loss
  • Training: AdamW, lr=\(10^{-4}\), weight decay=\(10^{-4}\), lr decay \(0.95\)/epoch; 15 epochs on 3DMatch, 30 on KITTI, 10 on nuScenes; single RTX 3090, batch size 1

Key Experimental Results

Dataset Metric Ours (IGASA) Prev. SOTA Gain
3DMatch (5000 pts) RR 94.6% SIRA-PCR 93.6% +1.0%
3DMatch (5000 pts) IR 87.9% RoITr 82.6% / SIRA-PCR 70.8% +5.3% / +17.1%
3DMatch (250 pts) RR 94.3% SIRA-PCR 92.4% +1.9%
3DLoMatch (5000 pts) RR 76.5% GeoTransformer 75.5% +1.0%
3DLoMatch (5000 pts) IR 61.6% RoITr 54.3% / SIRA-PCR 43.3% +7.3% / +18.3%
KITTI RR 100.0% Predator/GeoTransformer 99.8% +0.2%
KITTI RTE 4.6 cm OIF-Net 6.5 cm −1.9 cm
KITTI RRE 0.24° OIF-Net 0.23° on par
nuScenes RR 99.9% HRegNet 99.9% on par
nuScenes RTE 0.12 m HRegNet 0.18 m −33%
nuScenes RRE 0.21° HRegNet 0.45° −53%
Inference time (3DMatch) Total 2.763 s GeoTransformer 2.701 s +0.062 s

Ablation Study

  • Contribution of HCLA: baseline RR 89.6% → +HCLA 92.8% (+3.2%), validating the importance of cross-layer semantic alignment.
  • Contribution of IGAR: +HCLA 92.8% → +HCLA+IGAR 94.6% (+1.8%); IR increases from 79.2% to 87.9% (+8.7%), confirming that iterative geometric refinement is critical for outlier suppression.
  • SGIRA vs. SAIGA: SGIRA alone achieves FMR 96.2%; SAIGA alone achieves IR 84.2%; using both yields FMR 98.2% and IR 87.9%—demonstrating the synergy between semantic filtering and geometric sharpening.
  • Loss functions: Each individual loss term performs poorly (IR < 75%); joint optimization of all three achieves IR 87.9%, confirming that multi-task supervision is indispensable.
  • Computational efficiency: Total inference time of 2.763 s is highly competitive with GeoTransformer (2.701 s) and CoFiNet (2.660 s); the additional overhead of HCLA+IGAR is only ~0.1 s.

Highlights & Insights

  • IGAR's soft-suppression strategy replaces RANSAC's hard rejection, offering improved robustness and differentiability.
  • The dual-unit design of HCLA (semantic filter SGIRA + geometric sharpener SAIGA) effectively bridges the semantic gap.
  • IR improvements are substantial (87.9% on 3DMatch, 61.6% on 3DLoMatch), indicating a significant gain in correspondence quality.
  • When the number of sampled points drops from 5000 to 250, RR decreases only marginally from 94.6% to 94.3%, demonstrating strong robustness.
  • IGASA achieves 100% RR on KITTI with RTE of only 4.6 cm.

Limitations & Future Work

  • FMR on 3DLoMatch (80.5%) falls short of RoITr (89.6%) and SIRA-PCR (88.8%), indicating room for improvement in feature robustness under sparse sampling.
  • Iterative refinement (\(N=5\)) incurs additional computational cost; fewer iterations may be necessary for real-time applications.
  • Adaptability to highly dynamic environments has not been validated.
  • Comparisons with recent diffusion-based registration methods (e.g., PointDifformer) under large-rotation scenarios are not sufficiently explored.
  • vs. GeoTransformer: GeoTransformer (2022) models global dependencies with geometric Transformers but employs a simple fusion strategy without cross-layer attention; its IR on 3DMatch ranges from 71.9% to 85.1%, compared to IGASA's 87.9%. Inference times are nearly identical (2.701 s vs. 2.763 s).
  • vs. RoITr: RoITr (2023) uses rotation-invariant Transformers, achieving IR of 82.6%–83.0%, which outperforms GeoTransformer but remains below IGASA's 87.9%. RoITr achieves higher FMR (89.6% vs. 82.1%), suggesting that IGASA's feature recall warrants further improvement.
  • vs. SIRA-PCR: SIRA-PCR (2023) adopts sim-to-real adaptation, achieving RR of 93.6% on 3DMatch (vs. IGASA's 94.6%) and 73.5% on 3DLoMatch (vs. IGASA's 76.5%); the performance gap is more pronounced in low-overlap scenarios.

The design principle of SGIRA—using deep-layer semantics to guide shallow-layer feature fusion—is transferable to skip-connection optimization in U-Net-style architectures for medical image segmentation. IGAR's dynamic geometric consistency weighting is applicable to other outlier-rejection tasks such as visual localization and Structure-from-Motion. The multi-scale pyramid combined with cross-layer attention paradigm parallels EPT (another paper from the same batch), both emphasizing that information at different granularities requires differentiated processing.

Rating

  • Novelty: ⭐⭐⭐⭐ The dual-unit HCLA design and IGAR soft-suppression strategy are genuinely novel, though the overall approach remains within the coarse-to-fine paradigm.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Four datasets (3DMatch, 3DLoMatch, KITTI, nuScenes) with comprehensive ablation and runtime analysis.
  • Writing Quality: ⭐⭐⭐ Content is thorough but some formulas and notation are redundant; the related work section cites several papers with limited relevance to the core topic.
  • Value: ⭐⭐⭐⭐ Advances the state of the art across the board in point cloud registration, with particularly significant IR gains; open-source code facilitates broader adoption.
  • Novelty: ⭐⭐⭐
  • Experimental Thoroughness: ⭐⭐⭐
  • Writing Quality: ⭐⭐⭐
  • Value to Me: ⭐⭐⭐