检索结果-内蒙古大学图书馆

point cloud instance segmentation With Semi-Supervised Bounding-Box Mining

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 2022年第12期44卷 10159-10170页

作者： Liao, Yongbin Zhu, Hongyuan Zhang, Yanggang Ye, Chuangguan Chen, Tao Fan, Jiayuan Fudan Univ Sch Informat Sci & Technol Embedded Deep Learning & Visual Anal Grp Shanghai 200433 Peoples R China Agcy Sci Technol & Res Singapore 138632 Singapore Fudan Univ Acad Engn & Technol Shanghai 200433 Peoples R China

point cloud instance segmentation has achieved huge progress with the emergence of deep learning. However, these methods are usually data-hungry with expensive and time-consuming dense point cloud annotations. To alleviate the annotation cost, unlabeled or weakly labeled data is still less explored in the task. In this paper, we introduce the first semi-supervised point cloud instance segmentation framework (SPIB) using both labeled and unlabelled bounding boxes as supervision. To be specific, our SPIB architecture involves a two-stage learning procedure. For stage one, a bounding box proposal generation network is trained under a semi-supervised setting with perturbation consistency regularization (SPCR). The regularization works by enforcing an invariance of the bounding box predictions over different perturbations applied to the input point clouds, to provide self-supervision for network learning. For stage two, the bounding box proposals with SPCR are grouped into some subsets, and the instance masks are mined inside each subset with a novel semantic propagation module and a property consistency graph module. Moreover, we introduce a novel occupancy ratio guided refinement module to refine the instance masks. Extensive experiments on the challenging ScanNet v2 dataset demonstrate our method can achieve competitive performance compared with the recent fully-supervised methods.

关键词： point cloud compression Semantics Proposals Perturbation methods Three-dimensional displays Task analysis Annotations Semi-supervised learning weakly-supervised learning point cloud instance segmentation

来源：评论

学校读者我要写书评

暂无评论

point cloud instance segmentation with Inaccurate Bounding-Box Annotations

引用

SENSORS 2023年第4期23卷 2343页

作者： Peng, Yinyin Feng, Hui Chen, Tao Hu, Bo Fudan Univ Dept Elect Engn Shanghai 200433 Peoples R China Fudan Univ Yiwu Res Inst Yiwu 322000 Peoples R China

Most existing point cloud instance segmentation methods require accurate and dense point-level annotations, which are extremely laborious to collect. While incomplete and inexact supervision has been exploited to reduce labeling efforts, inaccurate supervision remains under-explored. This kind of supervision is almost inevitable in practice, especially in complex 3D point clouds, and it severely degrades the generalization performance of deep networks. To this end, we propose the first weakly supervised point cloud instance segmentation framework with inaccurate box-level labels. A novel self-distillation architecture is presented to boost the generalization ability while leveraging the cheap but noisy bounding-box annotations. Specifically, we employ consistency regularization to distill self-knowledge from data perturbation and historical predictions, which prevents the deep network from overfitting the noisy labels. Moreover, we progressively select reliable samples and correct their labels based on the historical consistency. Extensive experiments on the ScanNet-v2 dataset were used to validate the effectiveness and robustness of our method in dealing with inexact and inaccurate annotations.

关键词： point cloud instance segmentation learning with noisy labels weakly supervised learning self-distillation

来源：评论

学校读者我要写书评

暂无评论

Spherical Mask: Coarse-to-Fine 3D point cloud instance segmentation with Spherical Representation

Spherical Mask: Coarse-to-Fine 3D Point Cloud Instance Segme...

引用

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

作者： Shin, Sangyun Zhou, Kaichen Vankadari, Madhu Markham, Andrew Trigoni, Niki Univ Oxford Dept Comp Sci Oxford England

ISBN: (纸本)9798350353013;9798350353006

Coarse-to-fine 3D instance segmentation methods show weak performances compared to recent Grouping-based, Kernel-based and Transformer-based methods. We argue that this is due to two limitations: 1) instance size overestimation by axis-aligned bounding box(AABB) 2) False negative error accumulation from inaccurate box to the refinement phase. In this work, we introduce Spherical Mask, a novel coarse-to-fine approach based on spherical representation, overcoming those two limitations with several benefits. Specifically, our coarse detection estimates each instance with a 3D polygon using a center and radial distance predictions, which avoids excessive size estimation of AABB. To cut the error propagation in the existing coarse-to-fine approaches, we virtually migrate points based on the polygon, allowing all foreground points, including false negatives, to be refined. During inference, the proposal and point migration modules run in parallel and are assembled to form binary masks of instances. We also introduce two margin-based losses for the point migration to enforce corrections for the false positives/negatives and cohesion of foreground points, significantly improving the performance. Experimental results from three datasets, such as ScanNetV2, S3DIS, and STPLS3D, show that our proposed method outperforms existing works, demonstrating the effectiveness of the new instance representation with spherical coordinates. The code is available at: https://***/yunshin/SphericalMask

关键词： 3D instance segmentation instance segmentation point cloud instance segmentation

来源：评论

学校读者我要写书评

暂无评论

instance segmentation of point cloud captured by RGB-D sensor based on deep learning

引用

INTERNATIONAL JOURNAL OF COMPUTER INTEGRATED MANUFACTURING 2021年第9期34卷 950-963页

作者： Wang, Zhengtuo Xu, Yuetong Yu, Jiongyan Xu, Guanhua Fu, Jianzhong Gu, Tianyi Zhejiang Univ Inst Mfg Technol & Equipment Automat Coll Mech Engn State Key Lab Fluid Power & Mech Syst Hangzhou Peoples R China Zhejiang Univ Inst Mfg Technol & Equipment Automat Coll Mech Engn Key Lab Printing Proc & Equipment Zhejiang Prov 3 Hangzhou Peoples R China Suzhou Zijingang Intelligent Mfg Equipment Co Ltd Suzhou Peoples R China

RGB-D sensors are gradually introduced into the robotic system to help the machine understand its surroundings. Among the point cloud processing methods, instance segmentation of point cloud is extremely important since the quality of segmentation will affect the performance of subsequent algorithms. In this paper, the 3D reconstruction process of RGB-D sensor is analyzed, and a framework pointSeg is proposed to handle the instance segmentation of point cloud captured by RGB-D sensor. The pointSeg realizes point cloud instance segmentation by applying the deep learning method YOLACT++ to instance segment the color image first and then matching the instance information with the point cloud. In addition, an experimental platform that is equipped with a Kinect v2 is built, and a dataset is set up and then the pointSeg is tested on the dataset. The result shows that the pointSeg achieves point cloud instance segmentation according to the instance information extracted from color images, which not only has good real-time performance but also has better instance segmentation accuracy compared with the method of conducting instance segmentation directly on point cloud data, and the introduction of data augmentation in the training phase can achieve good training effect even on a small training dataset.

关键词： point cloud instance segmentation RGB-D sensor deep learning

来源：评论

学校读者我要写书评

暂无评论

Online segmentation of street trees from mobile laser scanning data via deep learning image instance segmentation

引用

OPTICS AND LASER TECHNOLOGY 2025年 184卷

作者： Li, Qiujie Gao, Junjie Nanjing Forestry Univ Coll Mech & Elect Engn Nanjing 210037 Peoples R China

A mobile laser scanning (MLS) system can efficiently collect the vertical structure information of street trees through light detection and ranging (LiDAR) movement. Segmenting street tree instances from street MLS data is the primary step in conducting street tree inventories at the single-tree scale, and existing methods are limited by the high computational complexity of three-dimensional (3D) point cloud processing, resulting in low speed and not supporting real-time applications. This paper proposes an efficient and effective online framework for segmenting street trees from street MLS data via deep learning image instance segmentation. First, the point cloud data collected in real time by LiDAR are stored in a first-in-first-out (FIFO) buffer. When the number of new LiDAR frames reaches the set segmentation step, the point cloud data in the FIFO buffer are converted into a three-channel image using point-pixel mapping. Next, a trained high-performance YOLOv8 segmentation model is used to segment an image instance of the street tree quickly. Then, a point cloud instance of the street tree is accurately segmented through optimization at the pixel and point levels. Finally, the segmentation step is adaptively adjusted on the basis of the street tree image instance segmentation result to improve the online processing efficiency. During the experiment, an MLS system equipped with a UTM-30LX-EW sensor is used to collect one-sided street point cloud data containing 149 street trees and other objects, such as landscape plants, buildings, viaducts, lanes, sidewalks, street lights, traffic signs, trash cans, baffles, fences, cars, motorcycles, and pedestrians. Eight YOLOv8 models with different target image sizes are compared, and the 416 x 320 model is selected to balance segmentation accuracy and speed, with an AP90 of 0.9950 and a segmentation time per image of 16.81 ms. Through pixel- and point-level optimization, the F1 score for online segmentation of street tree

关键词： Online segmentation Street tree Mobile laser scanning Image instance segmentation point cloud instance segmentation Adaptive segmentation step

来源：评论

学校读者我要写书评

暂无评论

3D rock fragmentation analysis using lidar, based on point cloud deep learning segmentation and synthetic data

引用

POWDER TECHNOLOGY 2025年 456卷

作者： Hafshejani, Mojgan Faramarzi Esmaeili, Kamran Univ Toronto Dept Civil & Mineral Engn 170 Coll St Toronto ON M5S 3E3 Canada

Accurate online measurement of particle size distribution is crucial in mining, tunnelling, and mineral processing industries to enable intelligent process control and optimization, ultimately enhancing efficiency and productivity. The current method for rock fragmentation relies on 2D image analysis, which is highly dependent on optimal lighting conditions, limiting its applicability and robustness in the challenging lighting environments commonly found in mining. This study diverges from the prevalent 2D image and photogrammetry approaches in rock fragmentation analysis, and pioneers a novel approach by harnessing laser scanner data for point cloud segmentation, offering a promising solution to overcome the limitations of image analysis techniques. By leveraging laser scanner data, a robust framework for rock fragmentation analysis is developed that is tailored to the specific challenges related to lighting situations. To avoid the laborious task of collecting and labelling point cloud datasets, this research introduces an innovative approach of using synthetic labeled datasets of scanned rockpiles. A platform is developed to automatically create and scan labeled point clouds of rock piles, facilitating the utilization of transfer learning. The synthetic 3D dataset was used to train a deep learning model for precise segmentation of rock instances in three-dimensional coordinates, providing an accurate representation of the rock object in 3D. The accuracy of the developed predictive model was tested and validated on experimental laser scanning data of three different rock piles. The proposed method depends on coordinate data instead of RGB information, rendering it particularly applicable in challenging conditions such as underground mining, night shifts, or situations where maintaining optimal lighting conditions is difficult or costly. The findings present a significant leap forward in rock fragmentation analysis, opening avenues for enhanced practices in diverse

关键词： Rock fragmentation analysis Deep learning point cloud instance segmentation Synthetic data generation 3D size distribution analysis Computer vision Laser scanner applications

来源：评论

学校读者我要写书评

暂无评论

Street tree segmentation from mobile laser scanning data using deep learning-based image instance segmentation

引用

URBAN FORESTRY & URBAN GREENING 2024年 92卷

作者： Li, Qiujie Yan, Yu Nanjing Forestry Univ Nanjing Peoples R China

Street tree inventory is an important part of urban forest inventory. Mobile laser scanning (MLS) technology has a strong data acquisition ability inside the canopy and the trunk and is suitable for parameter estimation at the tree level. The first and key step is to segment individual trees from the street MLS data, which is a challenging problem due to the semantic gap. This paper proposes an efficient and effective method for street tree segmentation from MLS data using deep learning -based image instance segmentation. First, the three-dimensional (3D) street point cloud captured by the MLS system is mapped to a two-dimensional (2D) RGB image. Next, pixelwise tree proposals are segmented from the street image by a trained deep learning -based image instance segmentation model. Then, the 2D segmentation mask is mapped back to the 3D street point cloud to generate pointwise tree proposals. Finally, the proposals are optimized to obtain the final results. To evaluate and verify algorithm performance, an MLS point cloud is collected from a 1481.8 m -long one -side urban street containing various objects. Three deep learning -based image instance segmentation algorithms, YOLACT, BlendMask, and YOLOv8, are carried out, and YOLOv8 achieves the best results in terms of both accuracy and speed. YOLOv8 has the highest segmentation accuracy, with IoU= 0.85:0.05:0.95, with an average segmentation time of 26 ms per image. In the experiments comparing the algorithms to the existing hierarchical segmentation and classification segmentation methods, the proposed method outperforms the other two methods in accuracy and is faster. The precision is 0.9988, the recall is 0.9986, the F1 score is 0.9987, and the time per scanline is 4.05 ms. Moreover, the proposed method can be applied to MLS data with a broad range of resolutions by introducing image resizing.

关键词： Street tree segmentation Mobile laser scanning Image instance segmentation point cloud instance segmentation Deep learning

来源：评论

学校读者我要写书评

暂无评论

Coarse-to-fine segmentation of individual street trees from side-view point clouds

引用

URBAN FORESTRY & URBAN GREENING 2023年 89卷

作者： Li, Qiujie Yan, Yu Li, Weizheng Nanjing Forestry Univ Nanjing Peoples R China

Segmenting individual street trees from a street side-view point cloud is the first and key step of obtaining a street tree inventory. Using the classification-segmentation framework for individual tree segmentation makes tree detection simple and accurate, but segmenting overlapping trees is still challenging. To more accurately segment overlapping trees, a coarse-to-fine method for segmenting individual street trees from a side-view point cloud is proposed in this paper. Following the classification-segmentation framework, the tree points are first detected from the side-view street point cloud by a pointwise classifier fused from 13 local geometric features and then trained using random forest (RF). Second, the tree proposals are obtained by density-based spatial clustering of applications with noise (DBSCAN) clustering and detection error filtering. Third, the overlapping tree proposals are recognized by trunk identification, and the single tree proposals are directly output as individual trees. Fourth, the overlapping trees are roughly divided into individual tree proposals through vertical planes. Finally, individual trees with optimized contours are obtained by iteratively using DBSCAN clustering and k-nearest neighbor (k-NN) classification. The side-view point cloud of a 290 m-long urban street containing 77 street trees is captured by a hand-held mobile ZEB Horizon laser scanner. The tree detection attained an F1 score of 0.9916 with a precision of 0.9989 and a recall of 0.9864. For individual tree segmentation, the F1 score was 0.9745 with a precision of 0.9672 and a recall of 0.9819. Compared to two current classification-segmentation methods, the overlapping tree segmentation F1 scores were increased by 0.0914 and 0.0617, respectively. The proposed method can be applied to tree parameter extraction, which is an important urban forest inventory task and is crucial for urban forest management. In our experiment, the root mean squared error (RMSE) of the tr

关键词： Individual tree segmentation Side-view point cloud Coarse-to-fine segmentation Overlapping tree segmentation point cloud instance segmentation

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：