How LEADER Beats Traditional LiDAR Relocalization in Accuracy and Speed

Problem Statement

LiDAR relocalization must estimate a vehicle’s 6‑DOF pose from a single point‑cloud fragment, e.g., when GPS is unavailable in underground garages.

Limitations of Existing Approaches

Retrieval‑registration pipelines achieve decimeter‑level accuracy but their storage and computation grow with map size. Neural‑network‑based methods—Absolute Pose Regression (APR) and Scene Coordinate Regression (SCR)—run in tens of milliseconds but are angle‑sensitive and typically only reach sub‑meter accuracy.

Research Question

Can SCR be made as accurate as retrieval‑registration while preserving its low storage and latency advantages?

LEADER Framework

LEADER (Learning Reliable Local‑to‑Global Correspondences for LiDAR Relocalization) addresses two dominant error sources in SCR:

Rotation sensitivity : Yaw changes can degrade accuracy from sub‑meter to >10 m.

Degenerate regions : Noisy or repetitive structures (long corridors, open floors) produce ambiguous point‑to‑world correspondences.

Key components:

Cylindrical projection + spatial transformation + Cartesian recovery : The point cloud is projected onto a cylinder and processed with cyclic sparse convolution, yielding yaw‑invariant features. Ground‑plane detection corrects roll and pitch, providing additional robustness.

TRR loss (Truncated Relative Reliability loss) : During training each point’s Euclidean distance loss serves as a proxy for predictability. The loss normalizes these proxies into confidence scores, which become per‑point weights in the total loss. Points with high loss receive lower weight, preventing over‑fitting to noisy data.

At inference, points with the highest predicted confidence are selected for RANSAC pose fitting, further reducing the influence of unreliable points.

Experimental Evaluation

Dataset: NCLT.

Median translation error:

APR: 1.19 m

SCR: 1.51 m

LEADER: 0.31 m

Comparison with rotation‑robust retrieval‑registration methods RING/RING++ (using the best of the two as reference):

Average XY error: LEADER 0.28 m vs. RING/RING++ (higher).

Failure rate within 5 m: LEADER 0.28 % (≈1/25 of RING/RING++).

Median error: LEADER 0.21 m, outperforming RING/RING++.

Confidence Analysis

Sorting test points by predicted confidence shows a clear negative correlation with pose error, confirming confidence as an effective filter. Using only high‑confidence points in the final RANSAC step improves accuracy further.

Ablation Study

Replacing TRR loss with standard Euclidean loss halves the proportion of high‑precision points, demonstrating that TRR enables the model to self‑adjust point weights and predict reliable confidences.

Insights

When full map memorization is unnecessary, allowing the model to select trustworthy points yields a simpler yet more effective solution. Redirecting model capacity toward reliable data can produce substantial performance gains.

Resources

Paper: https://arxiv.org/abs/2604.11355

Code repository: https://github.com/JiansW/LEADER