A Semantic Segmentation Network Exploiting Multiple Cues and Channel Exchanging

Semantic scene understanding is an important task in robotics and computer vision. In this paper, we propose an easily deployable and effective deep convolutional neural network structure for semantic segmentation of mobile robot platforms called Channel Exchange SegNet (CES). The network structure...

Full description

Saved in:
Bibliographic Details
Main Authors: Hao, Yici, Deng, Wenbang, Chen, Xieyuanli, Tang, Jingsheng, Lu, Huimin, Zhou, Zongtan
Format: Conference Proceeding
Language:English
Subjects:
Automation
Channel Exchanging
Computer vision
Convolutional neural networks
Information filters
Mobile robots
Multimodal Fusion
Normal Estimation
Semantic segmentation
Semantics
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Semantic scene understanding is an important task in robotics and computer vision. In this paper, we propose an easily deployable and effective deep convolutional neural network structure for semantic segmentation of mobile robot platforms called Channel Exchange SegNet (CES). The network structure is based on SegNet, a typical semantic segmentation network, and incorporates the Depth-to-Normal (D2N) module and the Channel Exchanging (CE) module. The D2N module utilizes the depth image to generate the normal image by three specially designed gradient filters. The CE module dynamically exchanges the feature maps between the three channels to achieve parameter-free multimodal fusion. We test our CES on the NYUDv2 and SUN-RGBD datasets. These quantitative evaluations show that our network has higher segmentation accuracy than single-mode networks.
ISSN:2688-0938
DOI:10.1109/CAC59555.2023.10450711