Adaptive Hierarchical Down-Sampling for Point Cloud Classification

Nezhadarya, Ehsan; Taghavi, Ehsan; Razani, Ryan; Liu, Bingbing; Luo, Jun

doi:10.1109/cvpr42600.2020.01297

Cited by 107 publications

(42 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In addition, Nezhadarya et al proposed a new deterministic, adaptive, and unchanging down-sampling layer called the critical point layer, which learns to reduce the number of points in the disordered point cloud while retaining the important (critical) point [41]. Unlike most graph-based point cloud down-sampling methods, the graph-based downsampling methods use K-nearest neighbor (K-NN) to find neighboring points.…”

Section: Reduction Methods Based On Deep Neural Networkmentioning

confidence: 99%

High-Precision 3D Reconstruction for Small-to-Medium-Sized Objects Utilizing Line-Structured Light Scanning: A Review

2021

View full text Add to dashboard Cite

Three-dimensional reconstruction technology has demonstrated broad application potential in the industrial, construction, medical, forestry, agricultural, and pastural sectors in the last few years. High-quality digital point cloud information exists to help researchers to understand objects and environments. However, current research mainly focuses on making adaptive adjustments to various scenarios and related issues in the application of this technology rather than looking for further improvements and enhancements based on technical principles. Meanwhile, a review of approaches, algorithms, and techniques for high-precision 3D reconstruction utilizing line-structured light scanning, which is analyzed from a deeper perspective of elementary details, is lacking. This paper takes the technological path as the logical sequence to provide a detailed summary of the latest development status of each key technology, which will serve potential users and new researchers in this field. The focus is placed on exploring studies reconstructing small-to-medium-sized objects, as opposed to performing large-scale reconstructions in the field.

show abstract

Section: Reduction Methods Based On Deep Neural Networkmentioning

confidence: 99%

High-Precision 3D Reconstruction for Small-to-Medium-Sized Objects Utilizing Line-Structured Light Scanning: A Review

2021

View full text Add to dashboard Cite

show abstract

“…First, it was necessary to perform point cloud clustering, segmentation, and noise suppression to extract the point cloud of the target. To increase the generalisability, the point cloud was rotated by a random angle around the z ‐axis (height direction) of the cluster centre and down‐sampled using the adaptive hierarchical down‐sampling method [30]. The point cloud was annotated based on the classification information recorded by the camera.…”

Section: Data Acquisitionmentioning

confidence: 99%

Traffic participants classification based on 3D radio detection and ranging point clouds

Bai

Tan

et al. 2021

IET Radar Sonar & Navi

View full text Add to dashboard Cite

Traffic participant classification is critical in autonomous driving perception. Millimetre wave radio detection and ranging (RADAR) is a cost‐effective and robust means of performing this task in adverse traffic scenarios such as inclement weather (e.g. fog, snow, and rain) and poor lighting conditions. Compared to commercial two‐dimensional RADAR, the new generation of three‐dimensional (3D) RADAR can obtain height information about targets as well as their dense point clouds, greatly improving target classification capabilities. This study proposes a multi‐objective classification method for traffic participants based on 3D RADAR point clouds. First, a 22‐dimensional feature vector of the 3D RADAR point cloud distribution was extracted to describe the shape, discrete, Doppler, and reflection intensity features of the targets. Then, dynamic and static datasets containing five classes of targets were produced, creating a 10k frame. Extensive experiments were conducted to build machine learning classifiers. The experimental results show that the trained classifiers can achieve over 92% classification accuracy when the targets are classified into five groups and over 95% classification accuracy when the targets are classified into four groups. The proposed method can guide the design of safer and more efficient intelligent driving systems.

show abstract

“…Recently, the deep learning-based simplification or sampling methods that are designed for special tasks such as classification, registration and recognition are becoming more popular. These methods, including FoldingNet [33], KCNet [34], SampleNet [35][36], PAT [37], CPL [38], MOPS-Net [39], PIE-NET [40], PointASNL [41], SK-Net [42], etc., attempt to sample the points with sensitive characteristics for effective feature learning. However, they do not achieve a good balance between uniform density and geometric feature keeping.…”

Section: Related Workmentioning

confidence: 99%