Multi-scale network optimization method based on Point Cloud Transform
The invention relates to a multi-scale optimization network method based on Point Cloud Transformer, and the multi-scale optimization network comprises three parts: the first part is a sampling layer which carries out the feature sampling of an input point cloud; the second part is of a multi-scale...
Saved in:
Main Authors | , , , |
---|---|
Format | Patent |
Language | Chinese English |
Published |
25.03.2022
|
Subjects | |
Online Access | Get full text |
Cover
Loading…
Summary: | The invention relates to a multi-scale optimization network method based on Point Cloud Transformer, and the multi-scale optimization network comprises three parts: the first part is a sampling layer which carries out the feature sampling of an input point cloud; the second part is of a multi-scale optimization structure and comprises a linear activation layer with convolution of different sizes; wherein the back of each convolution layer is connected with batchnorm and ReLU, linear activation layers with different convolution scales are used for carrying out composite feature extraction on sampled points, and then the sampled points are spliced with original input. The third part is a linear activation layer of the same-size convolution, the layer is mainly used for further feature extraction of data features output by the second part, and the function of the layer is similar to that of a full connection layer. Experimental comparison shows that the network structure constructed by the method is high in expe |
---|---|
Bibliography: | Application Number: CN202111563705 |