IGASA: Integrated Geometry-Aware and Skip-Attention Modules for Enhanced Point Cloud Registration¶
Conference: CVPR2026
arXiv: 2603.12719
Code: DongXu-Zhang/IGASA
Area: Autonomous Driving
Keywords: Point Cloud Registration, Geometry-Aware, Skip-Attention, Hierarchical Pyramid, Coarse-to-Fine Matching, Autonomous Driving
TL;DR¶
This paper proposes the IGASA framework, which employs a three-stage pipeline consisting of a Hierarchical Pyramid Architecture (HPA), Hierarchical Cross-Layer Attention (HCLA), and Iterative Geometry-Aware Refinement (IGAR) to bridge the semantic gap across multi-scale features and dynamically suppress outliers, achieving state-of-the-art performance on four benchmarks: 3DMatch, 3DLoMatch, KITTI, and nuScenes.
Background & Motivation¶
Point Cloud Registration (PCR) is a foundational task in 3D vision with direct applications in autonomous driving, robotic navigation, and environmental modeling. However, it remains challenging in real-world scenarios involving noise, occlusion, and large-scale transformations, where both accuracy and robustness are insufficient.
Traditional ICP-based methods rely on nearest-neighbor iterative minimization, making them sensitive to initialization and prone to local minima, with significant performance degradation under large misalignment or sparse data.
CNN-based methods (e.g., FCGF, D3Feat) are constrained by fixed receptive fields and struggle to model long-range dependencies. While Transformer-based methods (e.g., GeoTransformer, RoITr) can capture global context, fine geometric details are progressively diluted by aggressive downsampling as the network deepens—a phenomenon referred to as the "semantic gap."
Conventional skip connections typically employ naive fusion strategies such as concatenation or element-wise summation, which fail to properly calibrate the resolution mismatch between low-level geometric cues and high-level semantic embeddings, causing critical geometric details to be attenuated during fusion.
In the coarse-to-fine paradigm, the fine matching stage typically relies on RANSAC or hard-threshold pruning to remove outliers, which is computationally expensive and prone to discarding correct correspondences in low-overlap regions.
Consequently, a new framework is needed that simultaneously addresses the two key bottlenecks of multi-scale semantic alignment and robust outlier suppression.
Method¶
Overall Architecture¶
IGASA adopts a three-stage pipeline:
- HPA (Hierarchical Pyramid Architecture): Extracts features at three resolution levels (ordinary / minor / primary) using KPConv, with voxel sizes \(dl_0\), \(2 \cdot dl_0\), and \(4 \cdot dl_0\) respectively. The convolution radius is dynamically scaled to cover receptive fields ranging from local to global, producing \(F_{\text{multi}} = \{F_{\text{ordinary}}, F_{\text{minor}}, F_{\text{primary}}\}\).
- HCLA (Hierarchical Cross-Layer Attention): Serves as the core of coarse matching and comprises two sub-modules—SGIRA (Skip-Guided Inter-Resolution Attention) and SAIGA (Skip-Augmented Intrinsic Geometry Attention)—which explicitly align global semantics with local geometry to output $F_{\text{minor}}^}\(. Geometric consistency top-\)k$ selection then generates the coarse correspondence set \(\widetilde{C}^{(1)}\).
- IGAR (Iterative Geometry-Aware Refinement): During the fine matching stage, alternating optimization is performed via dynamic geometric consistency weighting, weighted centroid alignment, and SVD decomposition over \(N=5\) iterations to output the high-precision pose \(T^* = [R^*, t^*]\).
Key Designs¶
SGIRA Module: Uses global semantic features from the primary level as Query/Key to guide weighted fusion of high-resolution features from the minor level. The attention scores integrate three components:
- Semantic similarity \(S_{ij} = \frac{Q_i K_j^T}{\sqrt{d_a}}\)
- Geometric distance compensation \(R_{ij} = -\frac{\|P_i - M_j\|^2}{\sigma^2}\)
- Skip residual $F_{\text{minor}}^{ = F_{\text{minor}}^{+} + \gamma \cdot \text{SkipResidual}(F_{\text{minor}}^{+}, F_{\text{skip}})$
Fusion is realized through a Gated Fusion Mechanism: dual-branch convolution → adaptive gating weights → residual adjustment → weighted aggregation.
SAIGA Module: Applies self-attention on the SGIRA output \(F_{\text{minor}}^{+}\), integrating semantic similarity \(S_{\text{geo},ij}\) with learnable geometric distance weights \(R_{\text{geo},ij} = -\alpha \|M_i - M_j\|^2\), and introducing a skip attention bias \(\theta \cdot A_{\text{skip}}\). The output \(F_{\text{minor}}^{++}\) is obtained by aggregating the value matrix after Softmax normalization.
IGAR Module: At each iteration, correspondence weights are dynamically updated as \(w_{ij}^{(k)} = \exp\bigl(-\frac{\|p_{\text{tar}} - (R^{(k)} p_{\text{src}} + t^{(k)})\|^2}{\sigma^2}\bigr) \times \mathbb{I}[\cdot < \tau]\). Weighted centroids, a weighted cross-covariance matrix, and SVD decomposition are then used to solve for the optimal \(R^*, t^*\), constituting a soft-suppression rather than hard-pruning strategy for outlier handling.
Loss & Training¶
The total loss is \(\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{mat}} + \mathcal{L}_{\text{key}} + \mathcal{L}_{\text{den}}\), supervising three levels:
| Loss Term | Components | Role |
|---|---|---|
| \(\mathcal{L}_{\text{mat}}\) | Hierarchical matching probability loss \(\mathcal{L}_p\) + weighted cross-entropy \(\mathcal{L}_c\) | Supervises coarse matching probabilities |
| \(\mathcal{L}_{\text{key}}\) | InfoNCE descriptor loss \(\mathcal{L}_f\) + keypoint position loss \(\mathcal{L}_k\) + confidence BCE \(\mathcal{L}_i\) | Supervises keypoint matching |
| \(\mathcal{L}_{\text{den}}\) | Translation loss \(\mathcal{L}_t\) + rotation orthogonality constraint \(\mathcal{L}_r\) | Supervises global pose estimation |
Key Experimental Results¶
Indoor Benchmarks: 3DMatch & 3DLoMatch¶
| Method | 3DMatch RR(%) | 3DMatch IR(%) | 3DLoMatch RR(%) | 3DLoMatch IR(%) |
|---|---|---|---|---|
| GeoTransformer | 92.0 | 71.9 | 75.5 | 43.5 |
| RoITr | 91.9 | 82.6 | 74.7 | 54.3 |
| SIRA-PCR | 93.6 | 70.8 | 73.5 | 43.3 |
| IGASA | 94.6 | 87.9 | 76.5 | 61.6 |
- IGASA achieves the highest Registration Recall across all sampling rates (94.6%→94.3%), with minimal degradation as the number of samples decreases.
- The Inlier Ratio of 87.9% significantly outperforms RoITr (+5.3%) and SIRA-PCR (+17.1%).
Outdoor Benchmarks: KITTI & nuScenes¶
| Method | KITTI RTE(cm) | KITTI RRE(°) | KITTI RR(%) | nuScenes RTE(m) | nuScenes RR(%) |
|---|---|---|---|---|---|
| GeoTransformer | 6.8 | 0.24 | 99.8 | - | - |
| HRegNet | 12 | 0.29 | 99.7 | 0.18 | 99.9 |
| IGASA | 4.6 | 0.24 | 100.0 | 0.12 | 99.9 |
- IGASA achieves a 100.0% registration success rate on KITTI with an RTE of only 4.6 cm, the lowest among all compared methods.
- On nuScenes, RTE = 0.12 m and RRE = 0.21°, both state-of-the-art.
Ablation Study¶
| HPA | HCLA | IGAR | 3DMatch RR(%) | 3DMatch IR(%) |
|---|---|---|---|---|
| ✓ | - | - | 91.3 | 80.2 |
| ✓ | ✓ | - | 93.2 | 83.7 |
| ✓ | - | ✓ | 92.8 | 81.9 |
| ✓ | ✓ | ✓ | 94.6 | 87.9 |
Key Findings:
- The HCLA module contributes the largest RR improvement (+1.9%), validating the importance of cross-layer semantic alignment.
- The IGAR module contributes the most to IR improvement (83.7% → 87.9%), demonstrating that dynamic weight iteration effectively suppresses outliers.
- Joint training with all three loss terms is essential: using any single loss alone yields IR as low as 71–75%, while their combination achieves 87.9%.
- SGIRA and SAIGA exhibit synergistic effects: used independently, FMR is 96.2% and 96.7% respectively; combined, it reaches 98.2%.
- Inference speed is 2.763 s/frame, comparable to GeoTransformer (2.701 s) and CoFiNet (2.660 s).
Highlights & Insights¶
- Skip-Attention as a replacement for naive skip connections: SGIRA and SAIGA provide two levels of attention to bridge the multi-scale semantic gap, rather than resorting to simple concatenation or summation.
- Soft suppression instead of hard pruning: IGAR replaces RANSAC with a combination of dynamic geometric consistency weights and indicator functions, avoiding high computational overhead while reducing false rejection of correct correspondences in low-overlap regions.
- Comprehensive validation across four datasets: State-of-the-art results on both indoor (3DMatch/3DLoMatch) and outdoor (KITTI/nuScenes) benchmarks, with a notable 100% RR on KITTI.
- Clean modular design: The three stages—HPA → HCLA → IGAR—each serve a distinct role, with ablation studies thoroughly validating the necessity of each component.
Limitations & Future Work¶
- FMR on 3DLoMatch is not optimal: In low-overlap scenarios (10%–30%), the Feature Matching Recall (82.1%) falls below RoITr (89.6%) and GeoTransformer (88.3%), indicating that descriptor robustness under extremely low overlap remains improvable.
- Slight increase in inference latency: Approximately 0.1 s slower than CoFiNet, which may become a bottleneck in latency-critical applications.
- IGAR iteration count is a manual hyperparameter: \(N=5\) is determined empirically, with no adaptive termination mechanism.
- Evaluation limited to rigid registration: Applicability to non-rigid or dynamic scenes has not been explored.
- Training resource: A single RTX 3090 was used; scalability to large-scale continual training has not been discussed.
Related Work & Insights¶
- Traditional methods: ICP and its variants (sensitive to initialization, prone to local optima).
- CNN-based features: FCGF, D3Feat (fixed receptive fields, limited long-range dependency modeling).
- Transformer-based methods: GeoTransformer, RoITr, SIRA-PCR (strong global context but loss of fine-grained details).
- Coarse-to-fine frameworks: CoFiNet, PYRF-PCR (multi-scale fusion but reliance on RANSAC for fine matching).
- Skip connections: U-Net-style concatenation/summation (naive fusion leading to semantic gaps).
- The core distinction of IGASA lies in replacing naive skip fusion with attention mechanisms and substituting hard-threshold pruning with soft weighting.
Rating¶
- Novelty: ⭐⭐⭐⭐ — The idea of replacing naive skip connections with attention-based fusion is well-motivated; the soft-suppression mechanism in IGAR is soundly designed.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Four datasets, detailed ablations (modules / sub-modules / losses / efficiency), and qualitative visualizations.
- Writing Quality: ⭐⭐⭐⭐ — Clear structure with complete mathematical derivations; minor issues include symbol redundancy and an overly lengthy Related Work section.
- Value: ⭐⭐⭐⭐ — State-of-the-art on both indoor and outdoor benchmarks; 100% RR on KITTI demonstrates strong practical value; FMR on 3DLoMatch is not optimal.