MB-Loc: Multi-planar Bird's-eye-view Localization in outdoor LiDAR scenes

Abstract: Global LiDAR localization is a fundamental task for autonomous navigation systems. Recent methods perform Scene Coordinate Regression (SCR) and achieve superior accuracy over Absolute Pose Regression (APR) solutions by predicting dense 3D world coordinates. However, SCR approaches introduce two major bottlenecks: severe computational inefficiency from processing raw 3D geometries and significant performance degradation under varying sensor viewpoints. To address these limitations, we present MB-Loc, a lightweight and viewpoint-robust SCR framework. Instead of relying on heavy 3D convolutions, we project the input LiDAR scan into a 2.5D Multi-planar Bird's-Eye View (BEV) representation. By slicing the point-cloud along the Z-axis and mapping signed depths into discrete 2D planes, MB-Loc retains essential 3D geometric structures while exploiting the computational tractability of standard 2D CNNs. To handle the inherent sparsity of outdoor LiDAR, we introduce a KL-regularized latent bottleneck that explicitly models spatial uncertainty without injecting stochastic noise. Finally, to ensure rotation robustness, we apply 3D spatial augmentations prior to planar projection, forcing the network to implicitly learn viewpoint-invariant features. We perform extensive experiments on the publicly available NCLT dataset and demonstrate that our proposed method outperforms the current state-of-the-art. Operating at real-time inference speeds, MB-Loc significantly outperforms traditional 3D-SCR architectures in computational efficiency.

• The paper introduces a novel SCR pipeline using multi-planar BEV projection that enhances geometric fidelity and computational efficiency.
• It employs a deterministic latent bottleneck with KL regularization to robustly model uncertainty and mitigate overfitting on noisy LiDAR returns.
• Experimental results validate sub-meter accuracy and real-time inference on outdoor datasets, outperforming traditional APR and SCR methods.

Multi-planar Bird's-eye-view Localization in Outdoor LiDAR Scenes: An Expert Review

Motivation and Problem Statement

Global LiDAR localization is critical for precise 6-DoF pose estimation in autonomous navigation systems, underpinning tasks such as robot navigation and mapping within large, dynamic outdoor environments. Prevailing approaches include Absolute Pose Regression (APR), which regresses a pose directly from point-cloud data but suffers from ill-posedness due to reliance on memorization rather than explicit geometry, and Scene Coordinate Regression (SCR), which predicts dense 3D world coordinates for establishing correspondences via geometric registration. The latter, despite superior accuracy, incurs substantial computational overhead and is susceptible to performance degradation under viewpoint changes, primarily due to heavy 3D convolutional architectures and occlusion artifacts in 2D projections.

Method: MB-Loc Framework

MB-Loc introduces an SCR pipeline optimized for computational efficiency and geometric fidelity by projecting input LiDAR scans into a novel 2.5D Multi-planar Bird’s-Eye View (BEV) representation. The key architectural innovation involves partitioning the vertical Z-axis into multiple discrete planes, thus encoding spatial depth information across stacked 2D grids. By storing signed depths in each plane and employing scatter-reduce to resolve spatial collisions, MB-Loc exploits the efficiency of 2D CNNs while retaining critical structural features lost in conventional BEV projections. Dual spatial and channel attention mechanisms further enhance feature extraction.

To explicitly handle scene sparsity and uncertainty, MB-Loc integrates a KL-regularized deterministic latent bottleneck, inspired by Information Bottleneck principles. Rather than stochastic sampling, the network modulates feature maps via adaptively learned scale factors subject to a KL divergence penalty, regularizing structural uncertainty and mitigating overfitting to noisy returns. MB-Loc regresses 3D coordinate offsets for projected points, reconstructing dense 3D world coordinates for correspondence. A closed-form Singular Value Decomposition (SVD) RANSAC backend robustly estimates the global pose, rejecting outlier correspondences.

Figure 1: MB-Loc pipeline during inference, illustrating Z-axis slicing of the LiDAR point-cloud, multi-planar BEV projection, feature extraction, latent bottleneck regularization, and coordinate regression.

Multi-planar BEV Representation

MB-Loc’s spatial encoding projects the input point-cloud $\mathcal{P} = \{p_i \in \mathbb{R}^3\}$ into $P$ horizontal Z-slicing planes with spatial resolution $G \times G$. Each projected plane encodes local depth relative to its slice bounds, forming a sparse, multi-channel grid. Collisions are resolved by minimal relative depth retention, resulting in a primary tensor $\mathcal{V}$ for downstream processing. Additional auxiliary tensors include binary occupancy masks and local coordinate caches. By focusing regression on offset prediction within this bounded space, MB-Loc constrains learning to efficient local structural alignments.

Figure 2: Top: Stratified 3D point cloud partitioned along Z-axis; Bottom: Resultant 2D multi-channel BEV grid projections across $P=10$, $G=512$.

Network Design and Training

The encoder employs CBAM attention modules to isolate distinctive geometric features across Z-slicing planes. The deterministic latent bottleneck leverages parallel MLP heads to produce mean, variance, and scale maps, regularized via KL divergence toward a standard normal prior, bypassing stochastic sampling and enabling robust uncertainty modeling. Decoder layers inflate the latent state to native resolution, regressing per-plane 3D coordinate offsets. Training combines masked L1 regression (gated by occupancy masks) and a weighted KL penalty.

Pose Optimization and Backend

Dense 3D-3D correspondences from the network output provide input to a RANSAC-based SVD (Kabsch) solver for pose estimation. Iterative sampling and early termination optimize efficiency, while a dynamically tuned inlier threshold ($\tau$) balances outlier rejection and correspondence diversity.

Figure 3: Predicted (red) vs. ground truth (black) trajectories, showing MB-Loc’s alignment with reference paths.

Experimental Results

Accuracy-Latency Trade-off

MB-Loc achieves sub-meter localization on the NCLT dataset with a mean translation error of 0.91m and rotation error of 2.30°, outperforming baselines on translation accuracy and maintaining competitive rotation precision even across challenging test splits. The method’s robustness to seasonal variations is highlighted by its stable performance on the temporally distant 2012-05-26 sequence, whereas other baselines degrade sharply.

Figure 4: LiDAR localization performance vs. inference trade-off on NCLT, demonstrating MB-Loc’s optimal balance.

Uncertainty Evaluation

Pose uncertainty evaluation reveals that MB-Loc’s latent regularization yields lower variance and fewer error outliers compared to LightLoc, with stable error profiles correlated to predicted uncertainty.

Figure 5: Relationship between position error and variance on NCLT 2012-05-26; MB-Loc exhibits reduced outlier spikes.

Efficiency

MB-Loc completes forward inference in 6.8 ms and end-to-end pipeline (including pose resolution) in 21.1 ms, outperforming SCR baselines by up to $6.9\times$ in speed while using only 16M parameters—substantially fewer than SGLoc (105M) and DiffLoc (40M). These results confirm real-time deployment viability in computationally constrained environments.

Ablation Studies

Increasing Z-slicing planes and grid resolution improves accuracy by reducing quantization collisions and depth ambiguity. The best configuration ($P=15$, $G=512$) retains 31% of raw scan points but achieves optimal accuracy and runtime. Sensitivity analyses confirm that a threshold of $P$0 m and a correspondence subset size of $P$1 provide the best trade-off.

Figure 6: Reconstruction fidelity on a NCLT scan; MB-Loc’s refined cloud closely aligns with ground truth.

Trajectory Alignment

Qualitative trajectory plots across multiple NCLT sequences illustrate MB-Loc’s close adherence to ground-truth paths, with minimal drift or outlier deviations relative to APR and SCR baselines.

Figure 7: Predicted vs. ground-truth trajectories across NCLT sessions and baselines; MB-Loc demonstrates stable tracking.

External Dataset Generalization

On subsampled Oxford Radar RobotCar dataset, MB-Loc is competitive with state-of-the-art, though performance degrades with severe spatial sparsity, highlighting the impact of grid quantization and correspondence retention.

Figure 8: Trajectory comparison on subsampled Oxford dataset; MB-Loc and LightLoc attain strong results under extreme sparsity.

Limitations and Future Directions

MB-Loc’s fixed Z-axis slicing assumes relatively planar terrain and may underperform in highly non-planar environments. Severe dynamic occlusion is not addressed due to single-frame inference. Grid quantization in pre-subsampled point-clouds can reduce valid correspondences below robust registration thresholds. Future research should explore adaptive slicing, temporal scan fusion, and locally dynamic grid resolution to mitigate these effects.

Conclusion

MB-Loc establishes a new standard for efficient, robust SCR-based LiDAR localization in outdoor settings. By reconciling computational efficiency with geometric integrity via multi-planar BEV projection and deterministic latent regularization, MB-Loc successfully transcends the traditional accuracy-latency constraints of dense 3D localization pipelines. Its strong empirical results on the NCLT and Oxford datasets validate its practical deployment potential. The framework’s architectural and methodological innovations provide substantial impetus for subsequent research in uncertainty-aware localization, cross-domain generalization, and adaptive spatial representation, which are essential for advancing state-of-the-art autonomous navigation.