Source-linked AI summary
ReRadar: Robust Radar Global Localization via Rotation-Equivariant Descriptor Learning
Duc Manh Nguyen, Truong Giang Dao, Gia Nghiem Luong, Viet Trung Hoang, Anh Quang Nguyen
TL;DR
Many radar place-recognition descriptors discard spatial information needed for accurate pose retrieval. ReRadar learns equivariant features, builds rotation-invariant global descriptors, and combines retrieval with landmark matching for 3-DoF pose estimation. With target-dataset adaptation, it reports high Recall@1 on Mulran and Boreas, while cross-dataset adaptation also performs strongly on OORD.
Problem
Many radar place-recognition methods produce global descriptors without preserving the spatial information needed for accurate pose retrieval or global localization.
Source: §I · p. 1Method
ReRadar learns translation- and rotation-equivariant features, aggregates them with NetVLAD, and combines global retrieval with landmark-based matching for 3-DoF pose estimation.
Source: §I · p. 1Results
With target-dataset adaptation, ReRadar achieves 91.44% Recall@1 on DCC01 and reaches 99.38% Recall@1 on an evaluated Boreas sequence, surpassing the strongest external baseline on both evaluated Boreas sequences.
Source: B. Results on Mulran · p. 6, C. Results on Boreas · p. 7Takeaways & Limitations
Radar-domain adaptation transfers effectively across datasets, while target-dataset adaptation provides further gains on the most challenging sequences.
Source: §VI · p. 8
Abstract
from arXiv · showhide
Global localization with scanning millimeter-wave radar remains challenging because place-recognition descriptors often discard spatial structure needed for accurate pose retrieval. We present ReRadar, a radar global localization pipeline that extracts rotation-equivariant intermediate features using steerable convolutional neural networks, forms rotation-invariant descriptors through group pooling and NetVLAD aggregation, and combines descriptor retrieval with landmark-based matching to estimate the robot's three-degree-of-freedom (3-DoF) pose. Across fixed database-query evaluations, ReRadar with target-dataset adaptation achieves 99.37% Recall@1 on OORD Bellmouth, 91.44% Recall@1 with 80.99% F1_max on Mulran DCC01, and 99.38% Recall@1 on falling-snow Boreas sequence. Without target-dataset data, the cross-dataset model reaches 98.07% Recall@1 on OORD, performing comparably to the evaluated state-of-the-art methods.
I. INTRODUCTION
Radar supports robust long-range sensing in degraded conditions, but many radar place-recognition methods do not preserve the spatial information needed for pose retrieval. ReRadar combines equivariant feature learning, global retrieval, and landmark matching for 3-DoF localization.
I. INTRODUCTION
A. Radar-based Place Recognition
- Radar place-recognition approaches include handcrafted descriptors, classification, pretrained backbones with NetVLAD, and contrastive learning.Scanning-radar methods vary in representation and aggregation; several learning-based approaches target global recognition rather than local correspondence estimation.Source: §I · p. 1, A. Radar-based Place Recognition · p. 2
I. INTRODUCTION
- Most radar place-recognition methods produce global descriptors without preserving spatial information needed for accurate pose retrieval or global localization.This limits their direct applicability to tasks requiring accurate spatial correspondences.Source: §I · p. 1
- ReRadar learns translation- and rotation-equivariant features, aggregates them with NetVLAD, and uses landmark matching for final 3-DoF pose estimation.The introduction also presents a multi-stage descriptor-training strategy using InfoNCE [15] to improve robustness under cross-environment localization.Source: §I · p. 1
- ReRadar is evaluated on OORD, Mulran, and Boreas with fixed database–query configurations under in-dataset and cross-dataset training settings.On an NVIDIA Jetson Orin Nano, the reported end-to-end runtime is 294.98 ms per frame.Source: §I · p. 1
B. Radar Scan Matching
Radar scan-matching methods estimate relative motion from geometric or feature correspondences, while localization systems combine place recognition with metric pose estimation in different ways. ReRadar retrieves a candidate scan and then uses landmark matching with RANSAC to estimate its relative 3-DoF pose.
B. Radar Scan Matching
- Radar scan-matching methods establish geometric or feature correspondences, including radar-specific landmark detection and matching, graph matching, robust point-to-line registration, and decoupled rotation and translation estimation [1].These methods primarily align temporally or spatially nearby observations.Source: B. Radar Scan Matching · p. 2
- ReRadar retrieves a candidate radar scan, then applies radar-specific landmark matching and RANSAC-based 3-DoF pose estimation.This combines global place recognition with local registration after retrieval.Source: B. Radar Scan Matching · p. 2, C. Radar-based Place Recognition with… · p. 2
C. Radar-based Place Recognition with Pose Estimation
- Global localization requires both identifying a previously observed place and estimating the current observation's pose relative to the map.Radar-only localization avoids dependence on a pre-built LiDAR map, while some approaches estimate only initial heading rather than full pose.Source: C. Radar-based Place Recognition with… · p. 2
III. METHOD
ReRadar separates global place retrieval from relative pose estimation between the query and its retrieved database scan. It extracts global descriptors for retrieval and uses keypoints, local descriptors, and RANSAC to estimate the transformation.
III. METHOD
- The global localization task retrieves the most similar database entry, then estimates the query's relative pose to that entry.Database keyframe descriptors are precomputed, and nearest-neighbor search selects the best-matching index.Source: A. Global Localization Pipeline · p. 2
- A polar radar image is converted to Cartesian form for ReResNet [33] local descriptors, while a parallel landmark module detects keypoints; NetVLAD [11] aggregates features into a global descriptor.After global descriptors match, corresponding precomputed keypoint descriptors support relative-pose estimation via RANSAC [34].Source: Figure 1 · p. 3
- The pose-estimation module detects keypoints, assigns local descriptors, matches correspondences, and estimates a robust transformation with RANSAC.The estimated relative transformation contains translation and heading angle.Source: A. Global Localization Pipeline · pp. 2–3
1) Rotation-Equivariant and Invariant Descriptors:
ReRadar uses Cartesian radar images to retain translation-equivariant behavior, then learns rotation-equivariant intermediate features and forms a rotation-invariant global descriptor. Group pooling and NetVLAD provide the transition from spatial feature maps to a retrieval descriptor.
1) Rotation-Equivariant and Invariant Descriptors:
- Cartesian radar images retain the translation-equivariant behavior of standard CNNs, while steerable convolutions explicitly learn rotation-equivariant representations.The method uses a ResNet-style ReResNet encoder with steerable convolutions.Source: 1) Rotation-Equivariant and Invariant… · p. 3
- Group pooling removes the explicit orientation index to produce orientation-invariant channel vectors while retaining the feature map's spatial arrangement.The feature maps remain spatially equivariant.Source: 1) Rotation-Equivariant and Invariant… · p. 3
- NetVLAD aggregates flattened local descriptors using learnable cluster centers to form the global descriptor.Each k-th VLAD vector is computed before the vectors are combined into the descriptor.Source: 1) Rotation-Equivariant and Invariant… · p. 3
- Because NetVLAD aggregates spatial descriptors without preserving their ordering, the final global descriptor is rotation-invariant even though intermediate feature maps are equivariant.Under input rotation, feature-map positions rotate while their pooled-orientation channel representation is invariant.Source: 1) Rotation-Equivariant and Invariant… · p. 3
2) Pose Estimation:
ReRadar uses radar-specific keypoints and local descriptors to match a retrieved scan, then estimates the relative planar pose with RANSAC. The resulting translation and rotation form the 3-DoF pose estimate.
2) Pose Estimation:
- Radar-specific keypoints are more repeatable than traditional detectors on noisy, sparse radar scans, and the method adopts Cen et al. for its greater number of usable keypoints.SIFT, ORB, and FAST perform poorly under speckle noise and sparsity; the comparison is shown in Fig. 2.Source: 2) Pose Estimation · pp. 3–4, Figure 2 · p. 4
- Each keypoint indexes an orientation-invariant channel vector in the spatially equivariant feature map, and nearest-neighbor matching in descriptor space establishes tentative correspondences.Source: 2) Pose Estimation · p. 4
- RANSAC estimates the planar rigid transform by maximizing the number of inliers, yielding translation in meters and rotation in the metric Cartesian BEV frame.Pixel coordinates are converted to meters using the radar-image resolution of 0.64 m/pixel before RANSAC.Source: 2) Pose Estimation · p. 4
B. Network Training
ReRadar uses a two-stage training strategy to adapt its descriptor network, then optimizes InfoNCE so queries are closer to positive than negative descriptors.
1) Training Strategy:
- The two-stage strategy initializes ReResNet with ImageNet-pretrained weights, adapts it on KITTI sequence 00 BEV images, then fine-tunes it on radar data.Both stages use NetVLAD-aggregated global descriptors and InfoNCE with sampled positive and negative pairs.Source: 1) Training Strategy · p. 4
2) Loss Function:
- InfoNCE uses each query descriptor, a positive selected within a predefined spatial radius, and negative descriptors sampled outside it; the objective favors higher positive-pair similarity.Descriptors are ℓ2-normalized, making inner products cosine similarity; temperature τ controls similarity concentration.Source: 2) Loss Function · p. 4
A. Dataset
The evaluation uses OORD, Mulran, and Boreas to cover off-road, urban, and long-term seasonal and weather conditions, with specified database-query splits and training configurations.
A. Dataset
- The evaluation uses OORD for off-road environments, Mulran for urban scenes, and Boreas for long-term seasonal and weather variation.Source: A. Dataset · p. 4
1) OORD:
- OORD evaluation uses five benchmark trajectory pairs across Bellmouth, Hydro, Maree, and Two Lochs, each traversed multiple times along the same routes.Source: 1) OORD · p. 4
2) Mulran:
- Mulran uses single-session database-query splits at scan 765 for DCC01 and scan 1420 for KAIST03, testing repetitive structures and dynamic environments.Source: 2) Mulran · p. 4
3) Boreas:
- Boreas spans 44 sequences and more than 350 km, with repeated and unseen routes; queries include snow-covered and falling-snow sequences against a sunny database sequence.The sequences are selected to evaluate long-term place recognition across seasonal and weather variation.Source: 3) Boreas · pp. 4–5, Figure 3 · p. 5
B. Radar-Domain Training Settings
- Training compares KITTI-only, non-target radar adaptation, and target-dataset adaptation; Two Lochs 2 also serves as the OORD evaluation database, making that setting map-adapted rather than scene-disjoint.Source: B. Radar-Domain Training Settings · p. 5
C. Implementation Details
- Implementation uses ReResNet-50 with E(2)-CNN cyclic group G8 and 512-channel feature maps, processing 200 × 200 Cartesian images at 0.64 m/pixel.Training uses InfoNCE with τ = 0.1, 14 negative samples per query, batch size 2, and 30 epochs; the stated GPU is an NVIDIA RTX 3090.Source: C. Implementation Details · p. 5
D. Evaluation Metrics
The evaluation reports retrieval, precision-recall, geometric-error, and localization-success measures, with success thresholds varying by dataset.
D. Evaluation Metrics
- Recall@1 counts a retrieval as correct when the top-ranked result is within 25 m on OORD or 20 m on Mulran and Boreas, divided by the total number of queries.Source: D. Evaluation Metrics · p. 5
- F1_max summarizes the precision-recall trade-off independently of a fixed descriptor-distance threshold.Precision and recall are evaluated at threshold δ.Source: D. Evaluation Metrics · p. 5
- Geometric accuracy uses mean translation error in meters and rotation error in degrees over successful cases, while localization success requires joint translation and rotation thresholds.The thresholds are (25 m, 10°) on OORD and (20 m, 10°) on Mulran and Boreas; ReFeree success instead requires correct retrieval and rotation error below 10°.Source: D. Evaluation Metrics · p. 5
V. RESULTS
Across OORD and Mulran, target-dataset adaptation gives ReRadar its strongest retrieval results, while adaptation effects on pose metrics vary. Cross-dataset adaptation also improves selected routes without using target-dataset data.
V. RESULTS
- All three ReRadar settings use the same database–query pairs, enabling direct comparison of KITTI-only, cross-dataset, and target-dataset adaptation.Table I presents the quantitative OORD comparison.Source: §V · p. 5, Table I · p. 6
A. Results on OORD
- 98.07% Recall@1 on Hydro (1-vs-2) and 97.70% on Hydro (1-vs-3) for cross-dataset Our (2), despite no OORD adaptation data.KITTI-only Our (1) is competitive on most OORD pairs but declines on the more challenging Two Lochs route.Source: A. Results on OORD · p. 6
- 96.97% Recall@1 and 73% SR on Two Lochs for Our (3), up from 84.74% Recall@1 and 65% SR for Our (2); Our (3) leads external baselines in Recall@1 on all five OORD pairs.Our (3) has the highest F max 1 on four pairs, with RadVLAD marginally higher on Hydro (1-vs-3); rotation errors across equivariant settings are 2.27°–3.78°, and Two Lochs has the largest translation error.Source: A. Results on OORD · p. 6, Table I · p. 6
B. Results on Mulran
- 81.28% Recall@1 and 78.67% F max 1 on DCC01 for Our (2), improving on Our (1), while both settings remain competitive on KAIST03.On DCC01, Our (1) matches RadVLAD in Recall@1; the adaptation benefit is described as most pronounced on that route.Source: B. Results on Mulran · p. 6
- 91.44% Recall@1 on DCC01 and 98.33% Recall@1 on KAIST03: Our (3) leads both Mulran sequences, but its F max 1 is slightly below RadVLAD on KAIST03.Our (3) has the highest F max 1 on DCC01; its SR values are 88.24% and 97.99%, and the lowest rotation and translation errors occur in different ReRadar settings.Source: B. Results on Mulran · p. 6
C. Results on Boreas
On the two Boreas query sequences, target-dataset adaptation achieves high Recall@1 and supports accurate geometric pose estimation. The section also presents qualitative Mulran comparisons and descriptor-generation timing.
C. Results on Boreas
- 94.09% and 83.47% Recall@1: adaptation with Mulran data nearly matches RaPlace on one Boreas sequence, but a larger gap remains on the more challenging sequence.The KITTI-only model reaches 86.85% and 77.54% Recall@1 on the two sequences.Source: C. Results on Boreas · pp. 6–7
- Fig. 4 shows successful registration under rotation and translation; Figs. 5–6 compare trajectory-level retrieval consistency and same-query retrieval outputs across methods.Fig. 5 distinguishes true-positive and false-positive links by color; Fig. 6 uses a challenging Mulran KAIST03 query.Source: Figure 4 · p. 6, Figure 5 · p. 6, Figure 6 · p. 7
- 97.35% and 99.38% Recall@1: target-adapted ReRadar surpasses the strongest external baseline on both Boreas sequences.Its SR values are 97% and 99%; mean rotation errors remain below one degree and translation errors below 3.5 m.Source: Table III · p. 7, C. Results on Boreas · p. 7
D. Computational Time
- 90.40 ms on Jetson Orin Nano and 20.35 ms on RTX 3090: descriptor-generation runtime on Boreas; the end-to-end pipeline is close to real time at 4 Hz.ReRadar is faster than RaPlace and ReFeree but slower than RadVLAD and RingKey.Source: Figure 7 · p. 7, D. Computational Time · p. 7
E. Backbone Ablation
After target-dataset adaptation, ReRadar exceeds the strongest external baseline in Recall@1 on every evaluated OORD, Mulran, and Boreas pair. The backbone ablation and controlled-yaw evaluation show stronger retrieval robustness in challenging conditions and consistent rotation-equivariant local features across datasets.
E. Backbone Ablation
- 53.05, 5.34, and 6.39 percentage-point Recall@1 gains on OORD Two Lochs, Mulran DCC01, and Boreas 2020-11-26-13-58, respectively, show ReResNet's clearest cross-dataset gains on difficult shifts.On easier sequences, the backbones are closer.Source: E. Backbone Ablation · p. 7
- 4.51, 10.16, and 2.12-point Recall@1 gains on OORD Two Lochs, Mulran DCC01, and Boreas 2021-01-26-11-22, respectively, persist after target-dataset adaptation.Standard ResNet can match or slightly exceed ReResNet on some easier OORD pairs, while pose errors are generally comparable.Source: E. Backbone Ablation · p. 7
F. Controlled Yaw-Rotation Evaluation
- ReResNet maintains cosine similarity above approximately 0.95 on both datasets, while ResNet falls to approximately 0.84–0.87 at intermediate yaw rotations; the curves indicate transfer across datasets.The evaluation inverse-aligns feature maps from controlled yaw rotations before comparing their consistency.Source: F. Controlled Yaw-Rotation Evaluation · pp. 7–8, Figure 8 · p. 8
VI. CONCLUSION
- After target-dataset adaptation, ReRadar exceeds the strongest external baseline in Recall@1 on all evaluated OORD, Mulran, and Boreas database–query pairs.The conclusion also notes runtime overhead in landmark-based pose estimation as an area for improvement.Source: §VI · p. 8