Augmented-Training-Aware Bisenet for Real-Time Semantic Segmentation

Chih Chung Hsu, Cheih Lee, Shen Chieh Tai, Yun Zhong Jiang

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Semantic segmentation techniques have become an attractive research field for autonomous driving. However, it is well-known that the computational complexity of the conventional semantic segmentation is relatively high compared to other computer vision applications. Fast inference of the semantic segmentation for autonomous driving is highly desired. A lightweight convolutional neural network, the Bilateral segmentation network (BiSeNet), is adopted in this paper. However, the performance of the conventional BiSeNet is not so reliable that the model quantization could lead to an even worse result. Therefore, we proposed an augmented training strategy to significantly improve the semantic segmentation task's performance. First, heavy data augmentation, including CutOut, deformable distortion, and step-wise hard example mining, is used in the training phase to boost the performance of the feature representation learning. Second, the L1 and L2 norm regularization are also used in the model training to prevent the possible overfitting issue. Then, the post-quantization is performed on the TensorFlow-Lite model to significantly reduce the model size and computational complexity. The comprehensive experiments verified that the proposed method is effective and efficient for autonomous driving applications over other state-of-the-art methods.

Original languageEnglish
Title of host publicationICMEW 2022 - IEEE International Conference on Multimedia and Expo Workshops 2022, Proceedings
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9781665472180
DOIs
Publication statusPublished - 2022
Event2022 IEEE International Conference on Multimedia and Expo Workshops, ICMEW 2022 - Taipei City, Taiwan
Duration: 2022 Jul 182022 Jul 22

Publication series

NameICMEW 2022 - IEEE International Conference on Multimedia and Expo Workshops 2022, Proceedings

Conference

Conference2022 IEEE International Conference on Multimedia and Expo Workshops, ICMEW 2022
Country/TerritoryTaiwan
CityTaipei City
Period22-07-1822-07-22

All Science Journal Classification (ASJC) codes

  • Artificial Intelligence
  • Computer Science Applications
  • Computer Vision and Pattern Recognition
  • Media Technology

Fingerprint

Dive into the research topics of 'Augmented-Training-Aware Bisenet for Real-Time Semantic Segmentation'. Together they form a unique fingerprint.

Cite this