Multi-Scale Keypoint Detection Based on SHOT

Yongjie, 贾勇杰 Jia; Fengguang, 熊风光 Xiong; Xie, 韩燮 Han; Liqun, 况立群 Kuang

doi:10.3788/lop55.071013

Cited by 1 publication

(1 citation statement)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The point cloud is an irregular three-dimensional point set data that represents the surface information of an object and has been widely used in various fields such as autonomous driving [1], 3D reconstruction [2], vegetation monitoring [3], etc.Point cloud classification is the foundation of deep applications of point clouds, and classification work is mainly carried out in two directions: methods based on geometric features [4][5][6][7] and methods based on deep learning.Methods based on geometric features mainly use the geometric features of point clouds such as coordinates, curvature, and normal vectors to perform classification. This type of method is sensitive to noise and irregularities, and has poor universality.Methods based on deep learning use deep neural networks to extract feature representations of point cloud data.…”

Section: Introductionmentioning

confidence: 99%

A point cloud classification network with steadily increasing feature dimensions and attention mechanism

Wang,

Tian,

et al. 2023

International Conference on Remote Sensing, Mapping, and Geographic Systems (RSMG 2023)

View full text Add to dashboard Cite

To solve the problem of insufficient information extraction in deep learning point cloud classification networks due to the rapid increase in dimension, as well as the issue of channel information balance leading to a lack of better discrimination ability for object features, we propose a point cloud classification network with steadily increasing feature dimensions and attention mechanism. By transforming multiple combinations of channel numbers and layers, the optimal combination of channel numbers and layers in the PointNet feature extraction module was determined, and the channel numbers were changed from (64, 64, 64, 128, 1024) to (32, 64, 128, 256, 512, 1024), resulting in a network with a feature dimension that increases steadily by a factor of 2.Incorporate the SENet channel attention mechanism into the evenly dimensionality-increased network, which weights the feature channels to enable the network to identify more accurate features and thus improve the discriminative power of object features. After training and testing on the ModelNet40 dataset, the improved classification network achieved an overall accuracy improvement of 2.58% compared to the PointNet classification network.

show abstract