FlashReg: GPU-Accelerated 3-Clique Point Cloud Registration for Real-Time Correspondence-to-Pose Estimation

Ziyang Yu; Xiang Li; Qiong Chang; Jun Miyazaki

cs.CVAug 22, 2026

FlashReg: GPU-Accelerated 3-Clique Point Cloud Registration for Real-Time Correspondence-to-Pose Estimation

Authors: Ziyang Yu, Xiang Li, Qiong Chang, Jun Miyazaki

Organizations: School of Computing Institute of Science Tokyo Tokyo

Abstract

Graph-based point cloud registration achieves high robustness by identifying geometrically consistent correspondence sets, but constructing second-order compatibility graphs and enumerating candidate cliques remain compute- and memory-intensive. This work presents FlashReg, a GPU-oriented correspondence-to-pose estimator that avoids materializing the dense scored second-order graph. Its Fast First- and Second-Order Graph (FFSOG) construction builds a capacity-bounded sparse second-order graph directly from the binary first-order graph. A dataflow-optimized three-node clique (3-clique) search then selects pivots from compact per-row candidate pools and enumerates triples through sorted sparse-neighborhood intersections. Across indoor and outdoor benchmarks, FlashReg reduces correspondence-to-pose latency by 2--3x relative to TurboReg at comparable registration recall, while using about 50% of its peak allocated tensor memory on an embedded GPU. These results make FlashReg suitable as a high-throughput registration backend within onboard perception pipelines.

Figures & tables

Figure 1: Speed–accuracy trade-off on 3DMatch with FCGF descriptors; each label gives estimator throughput (FPS) and recall (%), and the truncated recall axis magnifies sub-point differences. FlashReg stays within half a point of the strongest evaluated solvers at nearly 200 FPS.

Figure 2: FlashReg pipeline. FastFOG and FastSOG form FFSOG, which emits column-sorted ELL-SOG rows and per-row top- $m$ candidate pools without materializing the dense SOG score matrix. The dataflow-optimized search consumes these outputs in place. Green and orange arrows denote on-chip and global-memory traffic, respectively.

Figure 3: FFSOG. Top, FastFOG: a fused, tiled kernel builds the FOG adjacency matrix and writes it to global memory once. Bottom, FastSOG: one block per node computes a row of the SOG score matrix in shared memory by common-neighbor voting and compacts the nonzeros via warp ballot and prefix sum; before the row leaves shared memory, its

m

highest-scoring retained edges are extracted into the compact pivot-candidate pool

\mathcal{A}_{i}

Device	Desktop PC	Jetson AGX Xavier
CPU	AMD Ryzen 7 9850X3D	NVIDIA Carmel
GPU	NVIDIA GeForce RTX 5090	512-core NVIDIA Volta
GPU memory	32 GB (VRAM)	16 GB (unified)
CUDA version	13.0	10.2

Table 1: Hardware and software configurations used in the experiments.

Learning-based
Method	FPFH				FCGF
Method	RR (%)	RE (deg)	TE (cm)	Time (s)	RR (%)	RE (deg)	TE (cm)	Time (s)
PointDSC*	77.51	2.07	6.45	0.035	92.91	2.04	6.52	0.034
VBReg*	79.54	2.27	7.01	0.072	93.04	2.07	6.57	0.079
Learning-free
RANSAC-10k	70.10	4.13	11.63	2.533	81.80	3.71	11.12	1.760

Table 2: Performance comparison of different registration methods on the 3DMatch dataset.

Learning-based
Method	FPFH				FCGF
Method	RR (%)	RE (deg)	TE (cm)	Time (s)	RR (%)	RE (deg)	TE (cm)	Time (s)
PointDSC*	27.51	4.24	10.42	0.032	55.14	3.85	10.45	0.033
VBReg*	33.58	4.38	10.77	0.074	57.94	3.80	10.75	0.074
Learning-free
RANSAC-10k	5.86	6.21	14.13	2.257	27.24	5.08	13.62	1.770

Table 3: Performance comparison of different registration methods on the 3DLoMatch dataset.

Learning-based
Method	FPFH				FCGF
Method	RR (%)	RE (deg)	TE (cm)	Time (s)	RR (%)	RE (deg)	TE (cm)	Time (s)
PointDSC*	97.29	0.44	9.56	0.047	98.74	0.21	11.01	0.046
VBReg*	97.83	0.43	9.48	0.054	98.92	0.28	10.92	0.055
Learning-free
RANSAC-10k	47.21	1.57	28.69	1.985	80.36	0.60	20.42	0.433

Table 4: Performance comparison of different registration methods on the KITTI dataset.

	Modules enabled					Per-stage GPU time (ms)
Configuration	FastFOG	FastSOG	Dataflow	Pivot	RR (%)	FOG	SOG	Pivot	Clique	Total
Dense reference					84.17	1.64	3.56	1.21	0.55	11.40
FastFOG	✓				84.17	0.19	3.56	1.21	0.55	9.95
FastSOG		✓			84.04	1.64	0.84	1.19	0.54	8.69
FastFOG $+$ FastSOG	✓	✓			84.04	0.19	0.84	1.19	0.54	7.22
FastSOG $+$ Dataflow (fused)		✓	✓	fused	84.01	1.64	0.42	–	0.04	6.55

Table 5: Component ablation on 3DMatch with FPFH. Total includes verification, refinement, and pose solving. Full (fused) is the deployed pipeline. All ablation experiments were conducted on the desktop PC (RTX 5090).

Figure 4: Recall, latency, total GPU-rail energy, and peak PyTorch-allocated tensor memory on the Jetson AGX Xavier (3DMatch) as

N_{\max}

and

K

vary. Energy is integrated over the complete 1,623-pair run. The “GPU Memory” panel denotes this allocator metric, not total device-memory use; the recall axis is truncated to expose small differences.

Point Cloud Registration

Point cloud registration aligns 3D point sets by estimating the transformations and correspondences that make their geometry consistent.

See all 111 papers

Robust Single-Parameter Point Cloud Registration using Truncated Lp Functions
Oct 25, 2026Seong Hun Lee, Patrick Vandewalle, Javier CiveraPoint Cloud Registration Robust Estimation
I3A, University of Zaragoza, 50018 Zaragoza, Spain · EAVISE, Department of Electrical Engineering, KU Leuven, 2860 Sint-Katelijne-Waver, Belgium
Generalized-CVO: Fast and Correspondence-Free Local Point Cloud Registration with Second Order Riemannian Optimization
Jun 8, 2026Ray Zhang, Marcus Greiff, Thomas Lew +1Riemannian Optimization LiDAR Odometry
Toyota Research Institute, 4440 El Camino Real, Los Altos, California, USA
PARTE: Plane-Assisted Robust Transformation Estimation for Point Cloud Registration
Sep 21, 2026Abolfazl Babanazari, Carson Cramer, Tyler Summers +2Point Cloud Registration
Department of Computer Science, Colorado School of Mines. · Department of Mechanical Engineering, University of Texas at Dallas. · U.S. Army Combat Capabilities Development Command, Army Research Laboratory

Oct 25, 2026J/K move · Enter open · S save

Robust Single-Parameter Point Cloud Registration using Truncated Lp Functions

Seong Hun Lee, Patrick Vandewalle, Javier Civera

I3A, University of Zaragoza, 50018 Zaragoza, Spain · EAVISE, Department of Electrical Engineering, KU Leuven, 2860 Sint-Katelijne-Waver, Belgium

Point Cloud Registration Robust Estimation

Open paper AskPDF arXiv

We propose a novel method for robust point cloud registration with only one tunable parameter. Our method, dubbed TLP, uses truncated

L_p

functions with different values of

p

to assign priority scores to correspondences, select the hypothesis, and refine the final model. By leveraging the robustness of truncated functions, TLP effectively handles the high outlier rates commonly encountered in real-world 3D feature matching. We also exploit GPU parallelization to evaluate hypotheses efficiently, enabling registration from thousands of correspondences in under a second on average. With only one tunable parameter, our method offers greater practicality than multi-parameter methods that require dataset-specific tuning of several parameters. Evaluations on public benchmark datasets demonstrate that TLP achieves higher accuracy than state-of-the-art single-parameter methods. Furthermore, it outperforms two widely used multi-parameter methods on three of the four evaluated datasets while maintaining competitive runtime.

Abstract

Point Cloud Registration

Explore similar work

Robust Single-Parameter Point Cloud Registration using Truncated Lp Functions

Generalized-CVO: Fast and Correspondence-Free Local Point Cloud Registration with Second Order Riemannian Optimization

PARTE: Plane-Assisted Robust Transformation Estimation for Point Cloud Registration