Point cloud completion is a critical task for intelligent vehicles in the Intelligent Transportation Systems (ITS). Due to the disordered nature of point clouds, current point-based point completion methods utilize max pooling operations and Convolutional Neural Networks (CNNs) with a convolution kernel size of 1 to extract point features. However, a network with a kernel size of 1 has a small receptive field and can only extract one point feature at a time; as a result, such a network struggles to precisely extract highly abstracted features. To address this issue, we propose a point cloud completion network with larger receptive fields, termed PLRF-Net, which is composed of two parts: a reconstruction network and a completion network. The reconstruction network focuses on repairing the structures of partial points and includes an offset-attention embedding module, a rearrangement module, several convolution layers with varying kernel sizes, a pooling and repeating layer, and transposed convolution layers with different kernel sizes. Meanwhile, the completion network strives to improve the precision of the completed point cloud and consists of several convolution layers, along with a downsampling and upsampling module. Experimental results indicate that our approach achieves superior quantitative and qualitative outcomes on both the Multiview Partial Point Cloud (MVP) and KITTI datasets.
Point Cloud Completion by Larger Receptive Fields for Intelligent Vehicles
2023-09-24
4885644 byte
Conference paper
Electronic Resource
English
Shaping receptive fields for affine invariance
IEEE | 2004
|Shaping Receptive Fields for Affine Invariance
British Library Conference Proceedings | 2004
|Road layout for larger vehicles
Engineering Index Backfile | 1964
|