Skip to main navigation Skip to search Skip to main content

DPA-PSMNet: Stereo Matching with Dynamic Pyramid Attention for Enhanced Depth Estimation

  • Lantao Li
  • , Chao Wei*
  • , Hongji Wang
  • , Ruijie Zhang
  • , Heying Huang
  • , Zhiqing Cao
  • *Corresponding author for this work
  • Beijing Institute of Technology
  • National Key Laboratory of Special Vehicle Design and Manufacturing Integration Technology

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Stereo image depth estimation plays a critical role in the field of autonomous driving. In most prior work, depth estimation results have often been unsatisfactory due to inadequate stereo matching accuracy. This work introduces an innovative architecture specifically designed to significantly improve stereo matching accuracy. Specifically, building upon the PSMNet framework, we introduce a Dynamic Pyramid Attention (DPA) mechanism to optimize and refine the feature extraction process. And the proposed approach undergoes comprehensive benchmarking on both Scene Flow and KITTI 2015 datasets. Empirical evidence confirms that our technique surpasses PSMNet in quantitative performance metrics.

Original languageEnglish
Title of host publicationProceedings of 2025 IEEE International Conference on Unmanned Systems, ICUS 2025
EditorsRong Song
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages780-786
Number of pages7
ISBN (Electronic)9798331526726
DOIs
Publication statusPublished - 2025
Externally publishedYes
Event2025 IEEE International Conference on Unmanned Systems, ICUS 2025 - Changzhou, China
Duration: 18 Sept 202519 Sept 2025

Publication series

NameProceedings of 2025 IEEE International Conference on Unmanned Systems, ICUS 2025

Conference

Conference2025 IEEE International Conference on Unmanned Systems, ICUS 2025
Country/TerritoryChina
CityChangzhou
Period18/09/2519/09/25

Keywords

  • DPA-PSMNet
  • dynamic pyramid attention
  • feature extraction
  • stereo matching

Fingerprint

Dive into the research topics of 'DPA-PSMNet: Stereo Matching with Dynamic Pyramid Attention for Enhanced Depth Estimation'. Together they form a unique fingerprint.

Cite this