DCReg (Decoupled Characterization for ill-conditioned Registration) is a principled framework for degenerate LiDAR registration. It decouples rotation and translation observability with Schur complements, maps eigenspaces into physical motion axes, and stabilizes only the weak directions through targeted preconditioning.

The main branch now contains the full public DCReg implementation. The earlier baseline-only public snapshot remains available on the separate baseline branch.

Highlights

• Schur-complement-based spectral degeneracy detection that removes misleading rotation-translation coupling before observability analysis.
• Physical-axis characterization that maps weak modes to roll/pitch/yaw and x/y/z, with aligned eigenvalues and contribution ratios.
• Targeted preconditioning that stabilizes only the weak directions instead of damping the full coupled system.
• Lightweight runtime stack: Eigen + PCL, with optional TBB/OpenMP.

Citation

@ARTICLE{hu2026dcreg,
   title={Dcreg: Decoupled characterization for efficient degenerate lidar registration}, 
   author={Xiangcheng Hu, Xieyuanli Chen, Mingkai Jia, Jin Wu, Ping Tan, Steven L. Waslander},
   journal={International Journal of Robotics Research (IJRR)}, 
   year={2026}
 }

News And Timeline

• 2026/05/26: Dcreg has been accepted by IJRR :tada:.
• 2026/04/27: opened upstream integration PRs for DCReg in Open3D and PCL, bringing degeneracy-aware point-to-plane registration ideas into widely used point-cloud libraries.
• 2026/04/21: released the verified full DCReg implementation on main.
• 2026/03/31: updated the second arXiv version and corrected the theoretical issue in the structured preconditioner analysis.
• 2026/03/12: received a Conditional Acceptance and started the final clarification and revision cycle.
• 2025/10/30: completed a Major Revision focused on clarifying the logic and presentation of the paper.
• 2025/09/23: released baseline codes and data, including ME-SR, ME-TSVD, ME-TReg, FCN-SR, O3D, XICP, and SuperLoc.
• 2025/09/09: released the preprint on arXiv.

Next Up

• Open-source a DCReg-based localization system to show how the method can be integrated into larger pipelines and adapted across different algorithms.

Quick Start

Dependencies

Tested on Ubuntu 20.04 with C++17.

Category	Packages
Required	Eigen3, PCL
Optional	TBB, OpenMP
Not required by C++ core	Ceres, yaml-cpp, Open3D

Build

From the repository root:

cmake -S DCReg -B DCReg/build
cmake --build DCReg/build -j8

Run

./DCReg/build/dcreg_minimal_example
./DCReg/build/dcreg_runner

Real-Scene Parking-Lot Demo

The repository also includes a compact real-scene runner:

cmake --build DCReg/build -j8 --target dcreg_parking_lot_example
./DCReg/build/dcreg_parking_lot_example

This runner keeps the core DCReg dependency surface unchanged: the C++ code still depends only on Eigen + PCL with optional TBB/OpenMP. The optional Open3D visualizer is a separate Python script:

python3 -m pip install open3d numpy pillow
python3 scripts/visualize_parking_lot_example.py

The visualization uses a black background, intensity-colored point clouds, a compact diagnostic card, and the detected weak/degenerate physical axes. For readability, the exported target map is a 100 m local crop of the prior map around the final pose, downsampled at 0.5 m while preserving intensity. In the bundled pk01-1976 frame, the current run converges in 5 iterations from the provided prior pose and characterizes the weakest translational direction along the physical x axis.

Bundled Sample Data

Small demo inputs are bundled under DCReg/data/ so the synthetic runner and parking-lot source frame are easy to inspect. The complete data package, including the large parking-lot prior map, is provided externally:

• Cylinder and parking-lot frames

For the parking-lot demo, download prior_map.pcd from that link and place it as:

DCReg/data/Parking-Lot-example/prior_map.pcd

What The Executables Are For

• dcreg_minimal_example: prints the three core DCReg modules and is the cleanest entry point for integrating the solver into another SLAM system.
• dcreg_runner: runs the verified synthetic registration pipeline and compares four parameterizations under the same implementation.
• dcreg_parking_lot_example: runs a real parking-lot single-frame-to-map registration case and exports visualization artifacts.

Representative Module-Level Log (dcreg_minimal_example)

dcreg_minimal_example is a synthetic linear-system example for showing the three DCReg modules. It does not run point-cloud registration, so it intentionally does not print RMSE, fitness, or runtime metrics.

Synthetic DCReg example
[Module 1] Spectral degeneracy detection
cond_full: 1122.345689
cond_schur_rot: 36.087707
cond_schur_trans: 742.066396

[Module 2] Physical-axis degeneracy characterization
degenerate_mask: 001001
raw_lambda_rot:  1.001796 19.881966 36.152505
raw_lambda_trans:  0.032394 12.252030 24.038714
aligned_lambda_rpy: 36.152505 19.881966  1.001796
aligned_lambda_xyz: 24.038714 12.252030  0.032394
rot_axis_contribution_ratio(each r_i as physical-axis mixture):
  r0 = 0.995659*roll + 0.000001*pitch + 0.004340*yaw
  r1 = 0.000004*roll + 0.999791*pitch + 0.000205*yaw
  r2 = 0.004337*roll + 0.000208*pitch + 0.995456*yaw
trans_axis_contribution_ratio(each t_i as physical-axis mixture):
  t0 = 0.999989*x + 0.000000*y + 0.000011*z
  t1 = 0.000000*x + 0.999834*y + 0.000166*z
  t2 = 0.000011*x + 0.000166*y + 0.999823*z
clamped_lambda_rpy: 36.152505 19.881966  3.615250
clamped_lambda_xyz: 24.038714 12.252030  2.403871

[Module 3] Preconditioned linear solve
preconditioned_delta:  0.190398 -0.306306 -2.620922  0.095321 -0.515479 41.997563
pcg_iterations: 6
pcg_relative_residual: 0.000000
qr_fallback: 0

Representative Registration Log (dcreg_parking_lot_example)

Registration metrics are emitted by dcreg_runner and dcreg_parking_lot_example. Runtime depends on the machine and parallel backend; the excerpt below shows the bundled parking-lot case on a local TBB run:

[Registration] Full real-scene matching
Test case: parking_lot_pk01
Algorithm: DCReg | Parameterization: SO3 | Parallel: TBB
Status: converged in 5 iterations
RMSE: 0.053225 | Fitness: 0.060365 | Time(ms): 1.86
Pose error: unavailable (no ground-truth pose for this case)
DCReg solver: pcg_iterations=6, relative_residual=0.000000, qr_fallback=0
Observability: schur_rot=1.957423, schur_trans=12.714237, mask=000100

Method Overview

DCReg is organized around three core modules:

1. Spectral degeneracy detection
   Build Schur complements for rotation and translation and inspect their spectra.
2. Physical-axis degeneracy characterization
   Align raw Schur eigenvectors to roll/pitch/yaw and x/y/z, then quantify contribution ratios and weak directions.
3. Preconditioned linear solve
   Clamp only the weak aligned eigenvalues and solve the normal equation with a targeted PCG update.

Schur-Based Detection	Physical-Axis Mapping

Verified Results On main

The default shifted_cylinder case was re-validated after the full public release sync:

Parameterization	Iter	RMSE	Fitness	Translation Error (m)	Rotation Error (deg)	LinIt	QRfb
Euler	9	0.0314988	0.116239	0.0258485	0.0516469	6	0
SE3	10	0.0315708	0.116636	0.0271197	0.0507196	6	0
SO3	10	0.0315708	0.116636	0.0271196	0.0507195	6	0
Quaternion	10	0.0315708	0.116636	0.0271196	0.0507195	6	0

Legacy Release Context

The earlier baseline-oriented public release remains available on the separate baseline branch.

Baseline Release Overview

Dataset Context	Release Context

Video Demo

Scenario	Characterization Example	Interpretation
		Planar degeneracy with dominant weak directions in X-Y-Yaw; see Fig. 16 in the paper.
		Sparse geometry in narrow stairs; weak directions move between t2 and r0-r1; see Fig. 17 in the paper.
		Narrow-passage degeneracy, typically r0-t0 or r0 depending on the measurements.
		Rich local structure inside geometrically narrow environments, often yielding r0-t0 or r0.

Experimental Results

Controlled Simulation Analysis

Overview

Error Trend	Convergence Trend

Real-World Performance Evaluation

Localization And Mapping

Localization	Mapping

Degeneracy Characterization

Characterization

Degeneracy Detection

Detection Case A	Detection Case B

Ablation And Hybrid Analysis

Ablation	Hybrid Analysis

Runtime Analysis

Runtime	Runtime Detail

Parameter Study

Technical Notes

Schur Conditioning

Figure	Interpretation
	S_R is the Hessian of the rotation subproblem after optimally accommodating translation, so its spectrum reflects rotation observability without translation-scale contamination.
	The Schur projection removes the component of range(J_R) that can be explained by J_t, preserving only irreducible rotation information.
	This explains why Schur complements naturally reduce sensitivity to unit and scale disparities between radians and meters.
	kappa(S_R) can differ substantially from kappa(H_RR) when cross-coupling is strong, which is exactly why coupled weak directions should be analyzed after decoupling.

Eigenvalue Clamping In Subspace

Figure	Interpretation
	DCReg clamps only the weak eigenvalues in the decoupled subspace instead of overwriting the full coupled Hessian.
	The same operation can also be interpreted as targeted regularization confined to the weak directions.

Documentation

Public-facing documentation lives on the GitHub Wiki.

Acknowledgment

The authors gratefully acknowledge the valuable contributions that made this work possible.

• We extend special thanks to Dr. Binqian Jiang and Dr. Jianhao Jiao for their insightful discussions that helped refine the theoretical framework of this work.
• We also appreciate Mr. Turcan Tuna for his technical assistance with the baseline XICP implementation.

Contributors