Source-linked AI summary
BEVPlace: Learning LiDAR-based Place Recognition using Bird's Eye View Images
Lun Luo, Shuhang Zheng, Yixuan Li, Yongzhi Fan, Beinan Yu, Siyuan Cao, Huiliang Shen
TL;DR
LiDAR place-recognition methods can lose performance under view variation or scene changes. This paper uses BEV images with a rotation-invariant network and relates feature distances to geometry distances for query-position estimation. Experiments report strong recall, generalization, and robustness to view changes, alongside position-estimation capability.
Problem
Existing LiDAR place-recognition methods using unordered points or range images may degrade under view variation or scene changes, and motion sensitivity can limit generalization.
Source: §1 · p. 2Method
BEVPlace uses BEV images, group convolution for rotation-equivariant local features, and NetVLAD for rotation-invariant global features, while a feature-to-geometry distance mapping supports query-position estimation.
Source: §1 · p. 2, §4.2 · p. 5Results
BEVPlace achieves high recall, strong generalization, and robustness to viewpoint changes, and estimates query-cloud positions.
Source: §6 · p. 8Takeaways & Limitations
The paper concludes that BEV images and rotation-invariant design support place recognition across viewpoint changes, while feature–geometry distance correlation extends the approach to position estimation.
Source: §6 · p. 8
Abstract
from arXiv · showhide
Place recognition is a key module for long-term SLAM systems. Current LiDAR-based place recognition methods usually use representations of point clouds such as unordered points or range images. These methods achieve high recall rates of retrieval, but their performance may degrade in the case of view variation or scene changes. In this work, we explore the potential of a different representation in place recognition, i.e. bird's eye view (BEV) images. We observe that the structural contents of BEV images are less influenced by rotations and translations of point clouds. We validate that, without any delicate design, a simple VGGNet trained on BEV images achieves comparable performance with the state-of-the-art place recognition methods in scenes of slight viewpoint changes. For more robust place recognition, we design a rotation-invariant network called BEVPlace. We use group convolution to extract rotation-equivariant local features from the images and NetVLAD for global feature aggregation. In addition, we observe that the distance between BEV features is correlated with the geometry distance of point clouds. Based on the observation, we develop a method to estimate the position of the query cloud, extending the usage of place recognition. The experiments conducted on large-scale public datasets show that our method 1) achieves state-of-the-art performance in terms of recall rates, 2) is robust to view changes, 3) shows strong generalization ability, and 4) can estimate the positions of query point clouds. Source codes are publicly available at https://github.com/zjuluolun/BEVPlace.
1. Introduction
BEV images offer a LiDAR representation less affected by sensor motion than point clouds or range images, motivating BEVPlace for robust place recognition. The paper also investigates estimating query-cloud positions from the relationship between feature and geometry distances.
1. Introduction
- Place recognition supports loop-closure constraints during map construction and relocalization when pose tracking is lost.LiDAR offers active sensing that is robust to illumination changes and precise depth information that can aid place recognition.Source: §1 · pp. 1–2
- Existing point-based and range-image methods are sensitive to LiDAR motion, and augmentation-based approaches risk overfitting and poor generalization.Range images can be distorted by point-cloud translations, while BEV object scale and position distributions remain nearly unchanged in the illustrated road scene.Source: Figure 1 · p. 1, §1 · p. 2
- BEVPlace uses group convolution to extract rotation-equivariant local features and NetVLAD to produce rotation-invariant global features.The authors report robustness to viewpoint changes and scene changes, with strong generalization and state-of-the-art recall on three large-scale datasets.Source: §1 · p. 2
- The paper explores the correlation between BEV feature distance and point-cloud geometry distance to estimate query-cloud positions.The authors describe this as the first position estimation directly from global descriptors, to their knowledge.Source: §1 · p. 2
- A simple BEV-based NetVLAD network outperforms state-of-the-art methods on KITTI and the benchmark dataset.The paper also reports that BEVPlace improves the baseline further.Source: Figure 1 · p. 1, §1 · p. 2
2. Related Work
The section organizes LiDAR place-recognition methods by whether they use 3D points or projection images. It reviews how point-based methods add local geometric or contextual features, while image-based approaches use different projections and matching strategies.
2. Related Work
- LiDAR place-recognition methods are grouped into approaches using 3D points and those using projection images as intermediate representations.The section points readers to [25] for a more comprehensive overview.Source: §2 · p. 2
- PointNetVLAD [1] and PCAN [36] treat points independently, while later methods add neighborhood, local-feature, or attention mechanisms.Examples include LPD-Net [19], EPC-Net [10], DH3D [6], SOE-Net [34], sparse-convolution MinkLoc3D [15] [16], and transformer-based methods [35] [37].Source: §2 · pp. 2–3
- Projection-image methods include range-image matching and overlap estimation, scan-context images, and rotation-invariant features from range images.BVMatch [20] uses handcrafted BEV features with bag-of-words, but is reported not to generalize well to unseen environments; the paper contrasts this with learned group-convolution features and NetVLAD [2].Source: §2 · p. 3
3. Preliminaries
Place recognition retrieves the database point cloud most structurally similar to a query, with structural similarity associated with geometric proximity. The paper maps point clouds to compact global features and extends retrieval with query-position estimation.
3. Preliminaries
- The database consists of point clouds paired with their sensor poses, and place recognition seeks the structurally similar database cloud for a query.In this setting, structural similarity is associated with collection at geometrically close places.Source: §3 · p. 3
- A learned mapping produces compact global features whose Euclidean distances rank structurally similar clouds ahead of dissimilar ones.Retrieval selects the database cloud with minimum feature distance to the query.Source: §3 · p. 3
- The paper trains on BEV images and extends place retrieval to estimate query-cloud positions.Source: §3 · p. 3
4. Method
BEVPlace maps point clouds to rotation-invariant global features from BEV images, then uses feature-to-geometry distance mappings to estimate query positions. Its design combines group convolution with NetVLAD and builds on an observed relationship between feature and geometry distances.
4.1. BEVPlace Network
- BEVPlace projects point clouds into BEV images and combines group convolution with NetVLAD to produce rotation-invariant global features.The BEV image uses normalized point density, with each ground-space grid measuring 0.4 m.Source: §4 · p. 3, §4.1 · p. 4
- Group convolution makes BEV features rotation-equivariant, and global pooling with NetVLAD yields rotation-invariant features.A rotated input transforms the feature map by a corresponding rotation before aggregation.Source: §4.1 · p. 4
- The network is trained with lazy triplet loss, which separates positive and negative feature distances by a margin; positives are defined by geometry distance below ϵ meters.The loss uses the hinge of the margin plus positive distance minus negative distance.Source: §4.1 · p. 4
4.2. Position Estimator
- Feature distance approximately increases with geometry distance and saturates for distant point clouds; the mean curves share a shape across KITTI sequences despite scene-dependent variation.The curves are modeled with a generalized Gaussian kernel.Source: §4.2 · p. 5
- A separate mapping is fitted for each database point cloud to recover query-to-reference geometry distances from feature distances.Each mapping is fit using that reference's feature and geometry distances to other database clouds.Source: §4.2 · p. 5
- The query position is estimated by minimizing squared errors between recovered distances and distances to nearby database references, using at least three references.The references are selected around the closest feature match.Source: §4 · p. 3, §4.2 · p. 5
- Minkloc3D-V2 and OverlapTransformer also have feature-to-geometry mappings approximately modeled by the same curve, supporting position estimation from their features.Their mapping shapes differ on KITTI sequences 00 and 06.Source: §4.2 · pp. 5–6
5. Experiments
Experiments compare BEVPlace across three public datasets, evaluating place retrieval, loop closure, position estimation, and runtime. BEVPlace shows stronger place-recognition performance and generalization, while its position estimator recovers distances and supports accurate localization.
5.1. Datasets
- Experiments used KITTI, ALITA, and benchmark data, with dataset-specific preprocessing and retrieval distance thresholds.KITTI frames were split into database and query sets; the benchmark point-cloud scale was unknown, but the method was reported to adapt to scale variation.Source: §5.1 · p. 6
5.2. Place Recognition
- On KITTI, BEV NetVLAD variants had comparable Recall at Top-1 to OverlapTransformer and outperformed unordered-point-based methods; BEVPlace achieved higher recalls than NetVLAD networks, validating the significance of rotation invariance.Source: §5.2 · pp. 6–7, Table 2 · p. 7
- On rotated KITTI, BEVPlace showed much higher recall and generalization than other methods, including rotation-invariant alternatives using points or range images.The test rotated point clouds around the z-axis to simulate view changes.Source: Table 3 · p. 7, §5.2 · p. 7
- On KITTI loop-closure curves, BEVPlace outperformed compared methods and generalized to other sequences despite training on part of sequence 00.The reported PR curves use a threshold on the Top-1 feature distance to determine positive matches.Source: §5.2 · p. 7, Figure 5 · p. 7
- BEVPlace outperformed other methods by large margins on the benchmark dataset and generalized well to ALITA after KITTI training.The benchmark comparison includes transformer-based methods; ALITA recall was evaluated on validation and test sets.Source: Table 4 · p. 7, §5.2 · p. 8, Table 5 · p. 8
5.3. Position Estimation
- Recovered geometry distances were more accurate, and BEVPlace and OverlapTransformer estimated positions more accurately than compared methods and remained robust under rotations.The distance-recovery evaluation used KITTI sequence 00; the rotation tests reported that other methods degraded substantially.Source: §5.3 · p. 8, Figure 6 · p. 8
5.4. Running Time
- Place recognition took about 30 ms per query and position estimation less than 1 ms, which the authors report supports real-time localization at 10 Hz.Runtime was measured on a desktop with an RTX3090 GPU.Source: §5.4 · p. 8
6. Conclusions
BEVPlace uses BEV images and a rotation-invariant network to achieve high recall, strong generalization, and robustness to viewpoint changes. The paper also relates feature distance to point-cloud geometry for position estimation and identifies 6-DoF pose estimation as future work.
6. Conclusions
- BEVPlace combines BEV images with a rotation-invariant network and achieves high recall, strong generalization, and robustness to viewpoint changes.The network is based on group convolution.Source: §6 · p. 8
- The observed correlation between geometry and feature distance supports a position-estimation model that can adapt to other place-recognition methods, with BEVPlace giving more accurate estimates.Source: §6 · p. 8
- Future work will encode rotation information into global features and estimate 6-DoF point-cloud poses.Source: §6 · p. 8