IGASA: Integrated Geometry-Aware Skip-Attention

• IGASA is a learning-based framework for rigid point cloud registration that integrates multi-scale KPConv-based feature extraction with geometry-aware skip-attention.
• It employs a Hierarchical Pyramid Architecture and Cross-Layer Attention (via SGIRA and SAIGA) to bridge the semantic gap between global context and local geometry.
• The Iterative Geometry-Aware Refinement module iteratively reweights coarse correspondences, improving pose estimation and achieving state-of-the-art performance on benchmarks.

IGASA, short for Integrated Geometry-Aware and Skip-Attention, is a learning-based framework for rigid point cloud registration (PCR) that combines a Hierarchical Pyramid Architecture (HPA), a Hierarchical Cross-Layer Attention (HCLA) module, and an Iterative Geometry-Aware Refinement (IGAR) module. It was introduced to improve registration robustness under heavy noise, significant occlusions, large-scale transformations, low overlap, and sparse outdoor LiDAR sampling, while preserving a coarse-to-fine correspondence-based pipeline and a weighted rigid alignment formulation (Zhang et al., 13 Mar 2026).

1. Registration setting and design objective

IGASA addresses rigid PCR between a source point cloud $P$ and a target point cloud $Q$. The task is to estimate a rigid transformation $T = (R, t)$, where $R \in SO(3)$ and $t \in \mathbb{R}^3$, such that the transformed source $RP + t$ aligns with the target. The paper formulates the source and target as

$$P = \{p_{\text{src},j}\} \subset \mathbb{R}^3,\quad Q = \{p_{\text{tar},j}\} \subset \mathbb{R}^3,$$

and writes the basic alignment objective over correspondences $\mathcal{C}$ as minimizing a squared Euclidean residual.

The practical pipeline remains correspondence based: detect keypoints, extract descriptors, match descriptors, and estimate $R,t$ with a robust solver. Within that setting, IGASA targets two bottlenecks identified in prior PCR systems. The first is the semantic gap created by hierarchical feature extraction: high-level, low-resolution features are semantically rich but geometrically coarse, whereas shallow features preserve local geometry but lack global context. The second is the dependence on hard outlier rejection, such as RANSAC or threshold-based pruning, during refinement. The framework is therefore organized so that HPA and HCLA improve multi-scale feature extraction and coarse matching, while IGAR replaces purely hard pruning with iterative geometry-aware weighting.

This placement situates IGASA within the family of modern learning-based PCR methods, but with an explicit emphasis on coupling multi-resolution feature fusion and downstream pose refinement rather than treating them as loosely connected stages.

2. Hierarchical Pyramid Architecture

The Hierarchical Pyramid Architecture is a three-level KPConv-based encoder that constructs multi-scale features for both source and target clouds. Its role is to provide a progression from fine local geometry to coarse global semantics.

For a query point $p_i$ with neighborhood $Q$0, KPConv aggregation is written as

$Q$1

with kernel weight function

$Q$2

where $Q$3 are learnable kernel points inside a sphere of radius $Q$4, and $Q$5 controls influence distance.

The hierarchy is built by grid subsampling with increasing voxel size $Q$6:

• Ordinary level: voxel size $Q$7, radius $Q$8, feature dimension 64.
• Minor level: voxel size $Q$9, feature dimension 128.
• Primary level: voxel size $T = (R, t)$0, radius $T = (R, t)$1, feature dimension 256.

The paper denotes the resulting feature set as

$T = (R, t)$2

The stated motivation is that increasing voxel size and receptive field coherently shifts the representation from local geometric fidelity to global semantic coherence. This is not treated as sufficient by itself: the hierarchy provides the substrate for later cross-scale alignment, but the paper argues that naive skip connections or simple concatenation do not resolve the semantic mismatch between levels.

3. Hierarchical Cross-Layer Attention and skip-attention

The Hierarchical Cross-Layer Attention module is the core mechanism used to bridge the semantic gap across scales. It consists of two submodules: Skip-Guided Inter-Resolution Attention (SGIRA) and Skip-Augmented Intrinsic Geometric Attention (SAIGA). Together they transform

$T = (R, t)$3

SGIRA

SGIRA uses coarse features to guide the refinement of mid-level features. Its inputs are $T = (R, t)$4, $T = (R, t)$5, and a skip feature $T = (R, t)$6. The paper describes a Gated Fusion Mechanism that normalizes coarse and mid-level features, passes them through parallel convolutional branches, computes gating weights, and applies residual correction to obtain a fused guidance feature.

Cross-layer similarity is written as

$T = (R, t)$7

and a geometry compensation term is added: $T = (R, t)$8 where $T = (R, t)$9 and $R \in SO(3)$0 are 3D coordinates from the two resolutions. Attention weights are formed from the Softmax over $R \in SO(3)$1 together with skip-derived weights, and aggregation yields

$R \in SO(3)$2

A skip residual correction then produces

$R \in SO(3)$3

This construction is intended to retain high-resolution detail while modulating it with globally informative features. The paper characterizes the effect as simultaneous semantic guidance, geometric consistency, and skip residual preservation.

SAIGA

SAIGA performs self-attention within the minor-resolution space while incorporating both geometric structure and skip information. Starting from $R \in SO(3)$4, it projects features into $R \in SO(3)$5, computes semantic similarity

$R \in SO(3)$6

adds geometric compensation

$R \in SO(3)$7

and defines attention as

$R \in SO(3)$8

Final refined features are then

$R \in SO(3)$9

The paper’s interpretation is that semantic similarity, geometric proximity, and skip augmentation together yield descriptors that are globally context-aware and locally precise. This suggests that the design is aimed not merely at better feature expressivity, but specifically at preserving correspondence reliability under large viewpoint changes and partial overlap.

Coarse superpoint matching

Using coarse features, IGASA forms superpoint correspondences by nearest-neighbor matching in feature space: $t \in \mathbb{R}^3$0

$t \in \mathbb{R}^3$1

This produces

$t \in \mathbb{R}^3$2

These matches are filtered by a geometric consistency score under an initial transform $t \in \mathbb{R}^3$3: $t \in \mathbb{R}^3$4 followed by a top-$t \in \mathbb{R}^3$5 selection to obtain $t \in \mathbb{R}^3$6, which is passed to IGAR.

4. Iterative Geometry-Aware Refinement

The Iterative Geometry-Aware Refinement module operates on the coarse correspondences and refines the transformation through iterative reweighting. Its defining property is that correspondence quality is updated using geometric consistency under the current transform, rather than relying only on one-shot hard rejection.

Initial weights are defined as

$t \in \mathbb{R}^3$7

At iteration $t \in \mathbb{R}^3$8, with current estimate $t \in \mathbb{R}^3$9, the weights become

$RP + t$0

The weighted alignment objective is

$RP + t$1

The solution is obtained by a weighted Procrustes procedure. Weighted centroids are

$RP + t$2

The weighted cross-covariance is

$RP + t$3

with

$RP + t$4

and the updated rigid transform is

$RP + t$5

IGAR repeats this cycle for a fixed number of iterations, set to $RP + t$6. In the paper’s terminology, the module is “geometry-aware” because weighting depends directly on residuals under the current alignment. A plausible implication is that IGAR occupies an intermediate position between purely learned correspondence scoring and purely classical robust estimation: coarse matching is learned, but refinement remains a principled weighted rigid solver.

5. Training strategy, datasets, and empirical performance

IGASA is trained with a composite loss

$RP + t$7

The matching loss supervises coarse correspondence probabilities across pyramid levels. The paper defines

$RP + t$8

and combines it with a weighted cross-entropy term $RP + t$9 into

$$P = \{p_{\text{src},j}\} \subset \mathbb{R}^3,\quad Q = \{p_{\text{tar},j}\} \subset \mathbb{R}^3,$$0

The keypoint matching loss supervises descriptors, positions, and confidence. Descriptor learning uses an InfoNCE objective $$P = \{p_{\text{src},j}\} \subset \mathbb{R}^3,\quad Q = \{p_{\text{tar},j}\} \subset \mathbb{R}^3,$$1, positional consistency is

$$P = \{p_{\text{src},j}\} \subset \mathbb{R}^3,\quad Q = \{p_{\text{tar},j}\} \subset \mathbb{R}^3,$$2

and confidence supervision is a binary cross-entropy term $$P = \{p_{\text{src},j}\} \subset \mathbb{R}^3,\quad Q = \{p_{\text{tar},j}\} \subset \mathbb{R}^3,$$3. These are combined as

$$P = \{p_{\text{src},j}\} \subset \mathbb{R}^3,\quad Q = \{p_{\text{tar},j}\} \subset \mathbb{R}^3,$$4

The dense registration loss supervises the final pose: $$P = \{p_{\text{src},j}\} \subset \mathbb{R}^3,\quad Q = \{p_{\text{tar},j}\} \subset \mathbb{R}^3,$$5 with

$$P = \{p_{\text{src},j}\} \subset \mathbb{R}^3,\quad Q = \{p_{\text{tar},j}\} \subset \mathbb{R}^3,$$6

Implementation details reported in the paper include PyTorch, HPA feature channels $$P = \{p_{\text{src},j}\} \subset \mathbb{R}^3,\quad Q = \{p_{\text{tar},j}\} \subset \mathbb{R}^3,$$7, 4-head attention with head dimension 64, one stack each of SGIRA and SAIGA, and AdamW with learning rate $$P = \{p_{\text{src},j}\} \subset \mathbb{R}^3,\quad Q = \{p_{\text{tar},j}\} \subset \mathbb{R}^3,$$8, weight decay $$P = \{p_{\text{src},j}\} \subset \mathbb{R}^3,\quad Q = \{p_{\text{tar},j}\} \subset \mathbb{R}^3,$$9, and learning-rate decay 0.95 per epoch. Voxel sizes are 0.025 m for 3DMatch/3DLoMatch and 0.3 m for KITTI/nuScenes. Data augmentation uses random rotations around arbitrary axes and random scaling. Training runs for 15 epochs on 3DMatch, 30 epochs on KITTI, and 10 epochs on nuScenes, with batch size 1 (Zhang et al., 13 Mar 2026).

The evaluation uses Inlier Ratio (IR), Feature Matching Recall (FMR), Registration Recall (RR), and on outdoor odometry benchmarks RRE and RTE. Reported results include the following.

On 3DMatch, IGASA achieves the highest RR across sample sizes from 250 to 5000 keypoints, at 94.3–94.6%, while maintaining IR = 87.9% across sample sizes. Its FMR is approximately 98.2%, although the paper notes that it is slightly below the best methods under very sparse sampling. On 3DLoMatch, IGASA reaches RR up to 76.5%, with IR around 61.5–61.6%, again exceeding several strong transformer baselines in registration accuracy.

On KITTI odometry, IGASA reports RTE = 4.6 cm, RRE = 0.24°, and RR = 100.0%. On nuScenes, it reports RTE = 0.12 m, RRE = 0.21°, and RR = 99.9%. The paper presents these values as the best RTE on KITTI, and the best RTE and RRE on nuScenes, with RR matching the top method on nuScenes.

Ablation studies attribute the gains to the interaction of HPA, HCLA, and IGAR. A baseline with only HPA yields RR = 91.3% and IR = 80.2% on 3DMatch. Adding HCLA increases RR to 93.2–93.4 with improved IR, and adding IGAR raises RR further to 94.6% with IR = 87.9%. A finer ablation reports that using neither SGIRA nor SAIGA gives FMR = 95.6, IR = 81.9, RR = 92.8; SGIRA only yields FMR = 96.2, IR = 83.3; SAIGA only yields FMR = 96.7, IR = 84.2; and using both reaches FMR = 98.2, IR = 87.9, RR = 94.6.

The computational profile on 3DMatch is reported as 0.135 s model time, 2.628 s pose time, and 2.763 s total time per pair. This is slightly slower than CoFiNet at 2.660 s and GeoTransformer at 2.701 s, but much faster than SpinNet at 88.152 s and faster than Predator at 7.064 s.

6. Context, applications, and limitations

IGASA is situated among classical registration approaches such as ICP, FGR, and RANSAC, as well as learning-based pipelines including FCGF, D3Feat, SpinNet, Predator, CoFiNet, GeoTransformer, RegTR, OIF-Net, SIRA-PCR, HRegNet, PointDSC, DGR, and EGST. Within that landscape, the paper identifies three main contributions (Zhang et al., 13 Mar 2026).

First, it introduces HPA integrated with cross-layer skip-attention (HCLA), replacing simple skip fusion with attention-driven, geometry-compensated multi-resolution interaction. Second, it proposes IGAR, an iterative refinement stage that uses soft, geometry-aware weighting of correspondences and a weighted Procrustes update rather than heavy RANSAC loops or purely threshold-based filtering. Third, it presents a coarse-to-fine design in which feature extraction and pose estimation are tightly coupled through explicitly geometrically consistent coarse matches.

The intended application domains are those in which robust rigid alignment is operationally important: autonomous driving, robotics, and environmental modeling. The reported KITTI and nuScenes results support use in LiDAR odometry, loop closure, and map building, while the indoor 3DMatch and 3DLoMatch results indicate applicability to scene reconstruction and general-purpose registration in cluttered environments.

The limitations explicitly stated in the paper are modest but consequential. IGASA incurs slight computational overhead because of the dual-unit HCLA design and the iterative IGAR stage. The paper also notes, as an implied limitation, dependence on supervised training with ground-truth poses and overlaps. Another plausible implication of the described three-level hierarchy is that extremely large-scale point clouds may require further scaling or partitioning. Future directions named in the paper include improved handling of highly dynamic environments and more efficient architectures for real-time, large-scale point cloud processing.

Code for IGASA is available at https://github.com/DongXu-Zhang/IGASA.