82914

Автор(ы): 

Автор(ов): 

2

Параметры публикации

Тип публикации: 

Доклад

Название: 

Height Estimation from Single Remote Sensing Images via Transfer Learning from Monocular Depth Estimation Models

ISBN/ISSN: 

2836-6131

DOI: 

10.1109/RusAutoCon65989.2025.11177401

Наименование конференции: 

  • 2025 International Russian Automation Conference (RusAutoCon)

Наименование источника: 

  • Proceedings of 2025 International Russian Automation Conference (RusAutoCon)

Город: 

  • Sochi

Издательство: 

  • IEEE

Год издания: 

2025

Страницы: 

725-731 https://ieeexplore.ieee.org/abstract/document/11177401
Аннотация
Height estimation from single remote sensing images remains a challenging problem due to limited labeled data and significant domain shifts from general-purpose imagery. This paper systematically explores the effectiveness of fully transferring pretrained monocular depth estimation models, particularly Depth Anything V2, for pixel-wise height prediction in aerial imagery. Our results demonstrate that retaining depth-specific decoders substantially enhances performance. Experiments on Data Fusion Contest 2018, ISPRS Potsdam, and ISPRS Vaihingen datasets show that Depth Anything V2, pretrained on 62M+ images, achieves a 7.2% MAE reduction over its backbone-only variant without depth-specific fine-tuning, and provides significantly sharper digital surface models. Our findings reveal that depth-pretrained models learn viewpoint-invariant geometric priors, enabling effective cross-domain transfer to height estimation tasks despite perspective shifts.

Библиографическая ссылка: 

Цуканов А.И., Чеканин В.А. Height Estimation from Single Remote Sensing Images via Transfer Learning from Monocular Depth Estimation Models / Proceedings of 2025 International Russian Automation Conference (RusAutoCon). Sochi: IEEE, 2025. С. 725-731 https://ieeexplore.ieee.org/abstract/document/11177401.