Abstract | ||
---|---|---|
In recent years, convolutional neural networks on point clouds have greatly improved the performance of point cloud classification and segmentation. However, the irregularity and disorder of point clouds make the convolution operation ill-suited to preserve the spatial-local structure and make the existing convolution networks very shallow. In order to solve the problems, we propose a novel pre-processor network named KeypointNet which ranks the point cloud according to the contribution to the final task, such as classification and segmentation. With the ordered point cloud, a convolution neural network on point clouds goes as deeper as possible similar to that on images. Two scoring mechanisms: directly scoring (DS) and gradually scoring (GS), are designed based on a pre-trained PointNet which is easily and fast trained. Both scoring mechanisms score a point depending on its contribution to the output of PointNet. The former scores the point cloud only one time and the later scores the point cloud on grade level. KeypointNet can be unified with any existing convolution neural networks on point clouds. Extensive experiments demonstrate that our method is effective and efficient and achieves SOTA classification results and comparable segmentation results, while greatly reducing space consumption. |
Year | DOI | Venue |
---|---|---|
2021 | 10.1007/978-3-030-87361-5_10 | IMAGE AND GRAPHICS (ICIG 2021), PT III |
Keywords | DocType | Volume |
Point cloud, Classification, Segmentation | Conference | 12890 |
ISSN | Citations | PageRank |
0302-9743 | 0 | 0.34 |
References | Authors | |
0 | 3 |
Name | Order | Citations | PageRank |
---|---|---|---|
Zuodong Gao | 1 | 0 | 1.01 |
Yanyun Qu | 2 | 216 | 38.66 |
Cui-Hua Li | 3 | 74 | 13.24 |