GLORN: Strong Generalization Fully Convolutional Network for Low-Overlap Point Cloud Registration

Xu, Jiabo; Huang, Yukun; Wan, Zeyun; Wei, Jingbo

doi:10.1109/tgrs.2022.3208380

Cited by 12 publications

(5 citation statements)

References 35 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…GCMTN is compared with 3DFeat-Net [53], FCGF [51], D3Feat [41], Predator [17] and GLORN [45], and evaluated on three metrics: RTE, RRE, and RR. The proportion of point cloud pairs where both RRE and RTE are below a certain threshold (RRE < 5•, RTE < 2 m).…”

Section: Registration Resultsmentioning

confidence: 99%

“…where o X and o Y denote the probability that the superpoints of the source and target point clouds lie in the overlap region, respectively. Inspired by GLORN [45], not all predicted points located in the overlap region are conducive to matching. Some superpoints may be located in smooth regions or corners, which will affect the generation of the best transformation parameters.…”

Section: Overlap Prediction Modulementioning

confidence: 99%

“…Regarding feature loss, relevant criteria have been established, wherein the feature similarity between matched point pairs is higher, whereas the feature similarity between unmatched point pairs is lower. To monitor the descriptors of pointwise features, some methods [17,41,45] employ circle loss [46] to extract the feature descriptors of superpoints via metric learning. But the circle loss ignores the differences between positive samples and weights them equally, so it is difficult to match the regions with relatively low overlap.…”

Section: Feature Lossmentioning

confidence: 99%

“…To investigate the characteristics of the proposed GCMTN, a comparison was also made between it and the related state-of-the-art algorithms, such as DCP-v2 [34], RPM-Net [33], Predator [17], REGTR [18], PerfectMatch [50], FCGF [51], D3Feat [41], SpinNet [52], 3DFeat-Net [53], and GLORN [45]. The parameters of these methods were adopted as the original papers reported.…”

Section: Implementation Detailsmentioning

confidence: 99%

See 3 more Smart Citations

GCMTN: Low-Overlap Point Cloud Registration Network Combining Dense Graph Convolution and Multilevel Interactive Transformer

Wang

Yang

2023

Remote Sensing

View full text Add to dashboard Cite

A single receptive field limits the expression of multilevel receptive field features in point cloud registration, leading to the pseudo-matching of objects with similar geometric structures in low-overlap scenes, which causes a significant degradation in registration performance. To handle this problem, a point cloud registration network that incorporates dense graph convolution and a mutilevel interaction Transformer (GCMTN) in pursuit of better registration performance in low-overlap scenes is proposed in this paper. In GCMTN, a dense graph feature aggregation module is designed for expanding the receptive field of points and fusing graph features at multiple scales. To make pointwise features more discriminative, a multilevel interaction Transformer module combining Multihead Offset Attention and Multihead Cross Attention is proposed to refine the internal features of the point cloud and perform feature interaction. To filter out the undesirable effects of outliers, an overlap prediction module containing overlap factor and matching factor is also proposed for determining the match ability of points and predicting the overlap region. The final rigid transformation parameters are generated based on the distribution of the overlap region. The proposed GCMTN was extensively verified on publicly available ModelNet and ModelLoNet, 3DMatch and 3DLoMatch, and odometryKITTI datasets and compared with recent methods. The experimental results demonstrate that GCMTN significantly improves the capability of feature extraction and achieves competitive registration performance in low-overlap scenes. Meanwhile, GCMTN has value and potential for application in practical remote sensing tasks.

show abstract

Section: Registration Resultsmentioning

confidence: 99%

Section: Overlap Prediction Modulementioning

confidence: 99%

Section: Feature Lossmentioning

confidence: 99%

Section: Implementation Detailsmentioning

confidence: 99%

See 2 more Smart Citations

GCMTN: Low-Overlap Point Cloud Registration Network Combining Dense Graph Convolution and Multilevel Interactive Transformer

Wang

Yang

2023

Remote Sensing

View full text Add to dashboard Cite

show abstract

“…Gümeli et al [38] proposed a neural network to predict pixellevel correspondences and utilize the Gauss-Newton optimization to improve the keypoint matching with those correspondences. For low overlapping situations, Xu et al [39] improve the performance by measuring the likelihood of the control points being in the overlap region and critical for matching. Li et al [40] decoupled pose estimation to point correspondence regression and pose estimation via this correspondence and proposed a novel inter-geometric consistency constraint loss to capture the point feature efficiently.…”

Section: Related Workmentioning

confidence: 99%

Online LiDAR-camera extrinsic parameters self-checking and recalibration

Wei,

Yan,

You

et al. 2024

Meas. Sci. Technol.

View full text Add to dashboard Cite

During production, smart cars are equipped with calibrated LiDARs and cameras. However, due to the vibrations that inevitably occur during driving, the sensors' extrinsic parameters may change slightly over time. It is a significant challenge to ensure the ongoing security of these systems throughout the car's lifetime. To address this issue, we propose a self-checking and recalibration algorithm that can continuously detect the sensor data of intelligent vehicles. If the sensor's miscalibration is detected, the data can be repaired promptly to ensure the vehicle's reliability.Our self-checking algorithm extracts features from the point cloud and image and performs pixel-wise comparisons. To improve feature quality, we utilize the patch-wise transformer to enhance local information exchange, which also benefits the subsequent extrinsic recalibration. To facilitate the study, we generate two customized datasets from the KITTI dataset and the Waymo Open Dataset. The experiments conducted on these datasets demonstrate the effectiveness of our proposed method in accurately calibrating the LiDAR and camera systems throughout the car's lifetime.This study is the first to highlight the importance of continually checking the calibrated extrinsic parameters for autonomous driving. Our findings contribute to the broader goal of improving safety and reliability in autonomous driving systems.The dataset and code are available at https://github.com/OpenCalib/LiDAR2camera_self-check.

show abstract

Addressing the generalization of 3D registration methods with a featureless baseline and an unbiased benchmark

Bojanić,

Bartol,

Forest

et al. 2024

Machine Vision and Applications

View full text Add to dashboard Cite

Recent 3D registration methods are mostly learning-based that either find correspondences in feature space and match them, or directly estimate the registration transformation from the given point cloud features. Therefore, these feature-based methods have difficulties with generalizing onto point clouds that differ substantially from their training data. This issue is not so apparent because of the problematic benchmark definitions that cannot provide any in-depth analysis and contain a bias toward similar data. Therefore, we propose a methodology to create a 3D registration benchmark, given a point cloud dataset, that provides a more informative evaluation of a method w.r.t. other benchmarks. Using this methodology, we create a novel FAUST-partial (FP) benchmark, based on the FAUST dataset, with several difficulty levels. The FP benchmark addresses the limitations of the current benchmarks: lack of data and parameter range variability, and allows to evaluate the strengths and weaknesses of a 3D registration method w.r.t. a single registration parameter. Using the new FP benchmark, we provide a thorough analysis of the current state-of-the-art methods and observe that the current method still struggle to generalize onto severely different out-of-sample data. Therefore, we propose a simple featureless traditional 3D registration baseline method based on the weighted cross-correlation between two given point clouds. Our method achieves strong results on current benchmarking datasets, outperforming most deep learning methods. Our source code is available on github.com/DavidBoja/exhaustive-grid-search.

show abstract

GLORN: Strong Generalization Fully Convolutional Network for Low-Overlap Point Cloud Registration

Cited by 12 publications

References 35 publications

GCMTN: Low-Overlap Point Cloud Registration Network Combining Dense Graph Convolution and Multilevel Interactive Transformer

GCMTN: Low-Overlap Point Cloud Registration Network Combining Dense Graph Convolution and Multilevel Interactive Transformer

Online LiDAR-camera extrinsic parameters self-checking and recalibration

Addressing the generalization of 3D registration methods with a featureless baseline and an unbiased benchmark

Contact Info

Product

Resources

About