Research map / 3D vision / Depth and stereo

SOAR: Regression-based LiDAR Relocalization for UAVs

NeurIPS 2026 · 3D vision › Depth and stereo · approach: Scene coordinate regression for localization · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SOAR, a scene-coordinate-regression LiDAR relocalization model for UAVs with constant-input serialization, sliding-window attention with yaw/altitude-invariant positional bias, and a softmax-free first layer, plus a new 4-scene UAVLoc dataset.

Problem

Map-free LiDAR relocalization methods designed for autonomous driving degrade sharply on UAVs due to arbitrary 6-DoF pose variations, yaw rotations, altitude changes, and sparse overlap from irregular flight paths.

Approach

Replaces XYZ input with constant features after voxel serialization (Hilbert/Morton), uses sliding-window local attention with a positional bias built from relative distance and pitch angle (yaw-invariant), and removes softmax in the first layer to avoid feature collapse under constant inputs; regresses per-point world coordinates followed by RANSAC pose estimation.

Key results

• On UAVLoc average, SOAR reaches 91.96% SR (5m,5°) and 1.12m median error, versus 48.90% SR for the best baseline RALoc.
• On UAVScenes AMtown_3 and AMvalley_3, SOAR achieves 91.57% and 88.82% SR (5m,5°) vs. 80.88% and 79.23% for SGLoc+RA.
• Ablation on AMvalley: removing the softmax-free first layer raises error from 2.14m to 212.13m mean error.

Surprising findings

• Using softmax in the first LoSWAtt layer causes complete failure (error above 100m), attributed to feature collapse under constant inputs.
• Rotation augmentation improved SGLoc on some UAVScenes sequences but reduced accuracy on HKairport (2.68m to 3.69m).

Limitations

• Requires per-scene training; a new model is needed for each environment.
• Theoretical invariance results are approximate, relying on an assumed bounded-perturbation condition.
• Dataset collected only in Xiamen, China, in four scene types.

Experimental setup

Datasets: UAVScenes UAVLoc

Benchmarks: UAVScenes UAVLoc

Metrics: Recall@1 (5m) success rate (5m,5°) RTE/RRE mean error median error

Baselines: SGLoc SGLoc+RA LightLoc RALoc HOTFormerLoc BEVPlace++ EgoNN Kiss-ICP PIN-SLAM

Builds on: Point Transformer v3 RANSAC SGLoc Hilbert curve Morton curve

Closest papers by meaning

• LightLoc: Learning Outdoor LiDAR Localization at Light Speed (CVPR 2025 · similarity 0.88)
• DiffLoc: Diffusion Model for Outdoor LiDAR Localization (CVPR 2024 · similarity 0.86)
• Game4Loc: A UAV Geo-Localization Benchmark from Game Data (AAAI 2025 · similarity 0.85)
• Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery (NeurIPS 2026 · similarity 0.85)
• OrthoLoC: UAV 6-DoF Localization and Calibration Using Orthographic Geodata (NeurIPS 2025 · similarity 0.85)
• Multi-Constellation-Inspired Single-Shot Global LiDAR Localization (AAAI 2024 · similarity 0.84)