LiDAR-based place recognition (LPR) is essential for global localization and loop-closure detection in large-scale SLAM systems. Existing methods typically construct global descriptors from Range Images or BEV representations for matching. BEV is widely adopted due to its explicit 2D spatial layout encoding and efficient retrieval. However, conventional BEV representations rely on simple statistical aggregation, which fails to capture fine-grained geometric structures, leading to performance degradation in complex or repetitive environments. To address this, we propose MPTF-Net, a novel multi-view multi-scale pyramid Transformer fusion network. Our core contribution is a multi-channel NDT-based BEV encoding that explicitly models local geometric complexity and intensity distributions via Normal Distribution Transform, providing a noise-resilient structural prior. To effectively integrate these features, we develop a customized pyramid Transformer module that captures cross-view interactive correlations between Range Image Views (RIV) and NDT-BEV at multiple spatial scales. Extensive experiments on the nuScenes, KITTI and NCLT datasets demonstrate that MPTF-Net achieves state-of-the-art performance, specifically attaining a Recall@1 of 96.31% on the nuScenes Boston split while maintaining an inference latency of only 10.02 ms, making it highly suitable for real-time autonomous unmanned systems.
Figures & tables
Fig. 1: Overview of the proposed MPTF-Net, a novel multi-view fusion-driven global descriptor extraction network for LiDAR-based place recognition.
Fig. 2: Overall pipeline of MPTF-Net. The network jointly exploits RIV and BEV representations containing geometric and intensity cues. RIV and BEV branches adopt ResNet-based backbones, and the multi-scale Transformer fusion module captures cross-view interactions. Finally, the context-gating enhanced NetVLAD aggregates the fused features into discriminative, viewpoint-invariant global descriptors.
Fig. 3: Block diagram of the BEV multi-feature encoding structure. After dividing the polar coordinate grids and selecting the point cloud clusters within the grids, NDT methods are utilized to compute geometric and intensity statistics.
Fig. 4: Visualization of multimodal BEV features. These maps capture complementary structural and radiometric information.
Approach
BS split
SON split
Recall@1
Recall@5
Recall@10
max F1
Recall@1
Recall@5
Recall@10
max F1
PointNetVLAD [ 4 ]
74.30
84.54
87.53
0.8885
98.49
99.43
99.58
0.9931
Scan Context [ 7 ]
85.71
93.27
97.07
0.9324
97.80
99.54
99.37
0.9941
MinkLoc3D [ 21 ]
89.37
94.82
96.15
0.9368
98.21
99.34
99.61
0.9917
FusionVLAD [ 16 ]
—89.21
95.61
98.10
0.9511
92.89
98.36
99.15
0.9891
LCPR [ 22 ]
94.15
98.44
99.14
0.9699
99.06
99.76
99.82
0.9953
TABLE I: Comprehensive evaluation of place recognition performance on the nuScenes dataset (BS and SON splits).
Approach
2012-02-05
2013-02-23
2013-04-05
Mean
R@1
R@5
R@20
R@1
R@5
R@20
R@1
R@5
R@20
R@1
R@5
R@20
PointNetVLAD [ 4 ]
0.746
0.823
0.875
0.469
0.604
0.719
0.449
0.576
0.683
0.555
0.668
0.759
Scan Context [ 7 ]
0.767
0.836
0.909
0.481
0.564
0.726
0.418
0.496
0.649
0.555
0.632
0.761
MinkLoc3D [ 21 ]
0.802
0.864
0.926
0.507
0.616
0.751
0.482
0.587
0.685
0.597
0.689
0.787
FusionVLAD [ 16 ]
0.786
0.870
0.922
0.510
0.643
0.754
0.429
0.553
0.667
0.575
0.689
0.781
LCPR [ 22 ]
0.856
0.891
0.935
0.669
0.798
0.830
0.651
0.712
0.802
0.725
0.800
0.856
TABLE II: Quantitative comparison of place recognition performance on the NCLT dataset. The best results are highlighted in bold , and the second-best results are underlined .
Dataset
Approach
AUC
max F1
KITTI
PointNetVLAD [ 4 ]
0.856
0.846
Scan Context [ 7 ]
0.836
0.835
MinkLoc3D [ 21 ]
0.894
0.869
FusionVLAD [ 16 ]
0.871
0.855
LCPR [ 22 ]
.862
0.854
CVTNet [ 17 ]
0.891
0.879
TABLE III: Quantitative comparison of place recognition performance on the KITTI dataset. The best results are highlighted in bold .
Approach
BS Split
R@1
R@5
R@10
Max F1
Ours (w/ NDT-BEV)
96.31
99.00
99.60
0.9813
w/ Std. BEV
94.21
98.40
99.00
0.9671
w/ Two-scale Fusion
93.61
94.01
97.14
0.9417
w/ Self-Attn
90.22
97.60
98.30
0.9353
TABLE IV: Ablation study of proposed MPTF-Net on the nuScenes dataset BS split.
Fig. 5: Analysis of multi-Scale fusion strategies on recall performance.
Fig. 6: Comparison of runtime and efficiency with state-of-the-art methods.
Fig. 7: Visual validation of rotation invariance. The upper left shows the original input; upper right displays the input after a 55 ∘ yaw rotation. The lower row presents the corresponding global descriptors generated by MPTF-Net, showing high structural consistency.
Fig. 8: Quantitative study on yaw-rotation invariance comparing Recall@1 across different rotation angles.