CADNet: Context-aggregated DCPPM monocular depth estimation network

Canjie Zhu, Huifang Sun, Mingfeng Lu*, Feng Zhang

*Corresponding author for this work

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

Abstract

Monocular depth estimation is a crucial technology for comprehending scenes, and acquiring global contextual information is pivotal for enhancing depth estimation accuracy. Traditional approaches for incorporating global context information involve pooling feature maps of varying receptive field sizes. Nevertheless, they fail to address challenges such as object boundary distortion and the loss of local detail information caused by complex textures and geometric structures in scenes. To tackle these issues, this paper proposes a novel monocular depth estimation model called CADNet (Context-aggregated DCPPM monocular depth estimation network). This model leverages a multi-scale context aggregation module, DCPPM, to effectively aggregate local features into a global framework, thereby resolving the problem of local detail loss during network training. Experimental results demonstrate that the CADNet model surpasses the NewCRFs model in complex scene boundary detection and capturing local object details. Furthermore, with a 6.27% reduction in parameter count, the CADNet model achieves a noteworthy 9.82% decrease in Sq Rel error on the KITTI dataset and exhibits remarkable performance in general depth estimation metrics for both indoor and outdoor scenes.

Original languageEnglish
Title of host publicationProceedings - 2024 International Conference on Electronic Engineering and Information Systems, EEISS 2024
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages161-165
Number of pages5
ISBN (Electronic)9798350351033
DOIs
Publication statusPublished - 2024
Event2024 International Conference on Electronic Engineering and Information Systems, EEISS 2024 - Changsha, China
Duration: 13 Jan 202415 Jan 2024

Publication series

NameProceedings - 2024 International Conference on Electronic Engineering and Information Systems, EEISS 2024

Conference

Conference2024 International Conference on Electronic Engineering and Information Systems, EEISS 2024
Country/TerritoryChina
CityChangsha
Period13/01/2415/01/24

Keywords

  • Contextual information aggregation
  • Depth boundaries
  • Fully connected conditional random field
  • Monocular depth estimation
  • Scene comprehension

Fingerprint

Dive into the research topics of 'CADNet: Context-aggregated DCPPM monocular depth estimation network'. Together they form a unique fingerprint.

Cite this

Zhu, C., Sun, H., Lu, M., & Zhang, F. (2024). CADNet: Context-aggregated DCPPM monocular depth estimation network. In Proceedings - 2024 International Conference on Electronic Engineering and Information Systems, EEISS 2024 (pp. 161-165). (Proceedings - 2024 International Conference on Electronic Engineering and Information Systems, EEISS 2024). Institute of Electrical and Electronics Engineers Inc.. https://doi.org/10.1109/EEISS62553.2024.00035