Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3-D Point Clouds

Abstract

Research on classification and segmentation of 3-D point clouds using deep learning methods has become a hot topic in emerging applications, such as autonomous driving, augmented reality, and indoor navigation. However, as the complexity of the network structures increases, the computational efficiency reduces, which affects the practical applications of these methods. In addition, prior researchers mostly seek to enhance the quality of spatial encodings, while the channel relationships are ignored. It makes the feature learning of point clouds insufficient, which will reduce the accuracy of classification and segmentation. In this article, a lightweight attention module (LAM) is proposed to improve the computational efficiency and accuracy at the same time by adopting a novel convolution mode and introducing a new attention mechanism based on channelwise statistical features. As the submodules of LAM, the lightweight module and the attention module can also be used independently to focus on improving the computational efficiency and accuracy, respectively, according to the actual applications. LAM and its submodules can be easily integrated into state-of-the-art deep learning methods on classification and segmentation of 3-D point clouds. The experimental results show that the proposed modules have a good performance on benchmark data sets.

3D Point Cloud Semantic Segmentation System Based on Lightweight FPConv

“…The recognition speed and portability of an NN are determined by model parameters and module sizes; specifically, a lighter module has a higher network speed and smaller module size but provides reduced overall network accuracy. The proposed model is compared with the networks proposed by other studies, which include PointNet [13], KPConv [22], FPConv [23], and the lightened models that were discussed in another study [32], namely SENet-PointNet, CBAM-PointNet, LAM-PointNet, SENet-PointNet++, CBAM-PointNet++, and LAM-PointNet++. The comparison results are presented in Tables VI.…”

“…In the comparison of the number of parameters, our proposed method is only 42% of PointNet [13], 10.1% of KPConv [22], 8.4% of FPConv [23], 36.6% of SENet-PointNet [32], 36.3% of CBAM-PointNet [32], 42.6% of LAM-PointNet [32], 74.2% of SENet-PointNet++ [32], 72.4% of CBAM-PointNet++ [32], and 87.5% of LAM-PointNet++ [32]. In addition, in the comparison of module size, our proposed method is only 60.44% of PointNet [13], 41.62% of KPConv [22], and 36.88% of FPConv [23].…”

“…The memory size of the PointNet [13], KPConv [22], and FPConv [23] methods require 14,000,000 bytes, 58,000,000 bytes, 69,720,000 bytes, and the access time is 23.3 msec, 96.7 msec, and 116.2 msec respectively. In addition, the memory size of SENet-PointNet [32], CBAM-PointNet [32], LAM-PointNet [32], SENet-PointNet++ [32], CBAM-PointNet++ [32], and LAM-PointNet++ [32] schemes requires 16,080,000 bytes, 16,200,000 bytes, 13,800,000 bytes, 7,920,000 bytes, 8,120,000 bytes, 6,720,000 bytes, the access time is 26.8 msec, 27 msec, 23 msec, 13.2 msec, 11.5 msec respectively. However, our proposed method only needs 5,880,000 bytes of memory size and the access time is 9.8 msec.…”

“…However, our proposed method only needs 5,880,000 bytes of memory size and the access time is 9.8 msec. In terms of memory usage of our proposed method, we only need 42% of PointNet [13], 10% of KPConv [22], 8% of FPConv [23], 37% of SENet-PointNet [32], CBAM -36% of PointNet [32], 43% of LAM-PointNet [32], 74% of SENet-PointNet++ [32], 72% of CBAM-PointNet++ [32], 88% of LAM-PointNet++ [32]. In terms of memory access time analysis, our method is 2.38 times faster than PointNet [13], 9.87 times faster than KPConv [22], 11.86 times faster than FPConv [23], and 2.73 times faster than SENet-PointNet [32].…”

“…In terms of memory usage of our proposed method, we only need 42% of PointNet [13], 10% of KPConv [22], 8% of FPConv [23], 37% of SENet-PointNet [32], CBAM -36% of PointNet [32], 43% of LAM-PointNet [32], 74% of SENet-PointNet++ [32], 72% of CBAM-PointNet++ [32], 88% of LAM-PointNet++ [32]. In terms of memory access time analysis, our method is 2.38 times faster than PointNet [13], 9.87 times faster than KPConv [22], 11.86 times faster than FPConv [23], and 2.73 times faster than SENet-PointNet [32]. 2.76 times faster than CBAM-PointNet [32], 2.35 times faster than LAM-PointNet [32], 1.35 times faster than SENet-PointNet++ [32], 1.38 times faster than CBAM-PointNet++ [32], and faster than LAM-PointNet++ [32 ] 1.14 times faster.…”