This paper presents a plug-in framework that enhances the robustness and cross-domain generalization of self-supervised monocular depth estimation models by leveraging unpaired real-world images and adversarial training, overcoming the limitations of paired-image dependency and domain-specific encoders for reliable deployment in diverse real-world conditions.

ULDepth: Transform Self-supervised Depth Estimation to Unpaired Multi-domain Learning
Phan Thi Huyen Thanh¹, Trung Thai Tran², The Hiep Nguyen², Minh Huy Vu Nguyen², Tran Vu Pham², Truong Vinh Truong Duy¹, Duc Dung Nguyen²

1 Aisin Corporation
2 Ho Chi Minh City University of Technology

Related link:
URL: https://ieeexplore.ieee.org/document/11122640
DOI: 10.1109/OJSP.2025.3597873