AMSASeg: An Attention-Based Multi-Scale Atrous Convolutional Neural Network for Real-Time Object Segmentation From 3D Point Cloud

Abstract

Extracting meaningful information on objects varying scale and shape is a challenging task while obtaining distinctive features on small to large size objects to enhance overall object segmentation accuracy from 3D point cloud. To handle this challenge, we propose an attention-based multi-scale atrous convolutional neural network (AMSASeg) for object segmentation from 3D point cloud. Specifically, a backbone network consists of three modules: distinctive atrous spatial pyramid pooling (DASPP), FireModule, and FireDeconv. The DASPP utilizes average pooling operations and atrous convolutions with different sizes to aggregate distinctive information on objects at multiple scales. The FireModule and FireDeconv are responsible to efficiently extract general features. Meanwhile, a spatial attention module (SAM) and channel attention module (CAM) aggregate spatial and semantic information on smaller objects from low-level and high-level layers, respectively. Our network enables to encode multi-scale information and extract distinct feature on overall objects to enhance segmentation performance. We evaluate our method on KITTI dataset. Experimental results demonstrate that the proposed network is effective to improve segmentation performance on small to large objects at real-time speed. INDEX TERMSDeep learning, Convolutional neural network, Object segmentation, 3D point cloud, Autonomous vehicles.

A token-mixer architecture for CAD-RADS classification of coronary stenosis on multiplanar reconstruction CT images

“…Further, to verify the hypothesis of effectiveness in modeling long-range dependencies, as the majority of CNNs adopt pyramid structure to compute multi-scale feature efficiently, we compared the proposed model (ConvMixer) with CNN-like pyramid architectures. Inspired by recent work in processing multi-scale information [52,53], in the tested CNNs conventional convolutions were replaced with either atrous convolutions (AtrousCNN) or Inception module [54] (InceptionV2CNN, InceptionACNN, Incep-tionBCNN). Specifically, an Inception-v2 version [26] was adopted for InceptionV2CNN, while an inspired Inception-ResNet version [46] was used in both InceptionACNN and InceptionBCNN.…”

PLU‐Net: Extraction of multiscale feature fusion

“…However, these methods typically use the original ASPP structure, which utilizes multiple branches to obtain rich feature information but also carries a significant amount of redundant information. The AMSASeg 16 method adds a regular convolution operation before each dilated convolution operation in each ASPP branch. The SAR‐U‐Net 17 method employs a three‐branch ASPP structure, added to the original five‐layer deep U‐Net network, while the SENetCount 18 method uses ASPP for skip connections to integrate low‐level information.…”