Hf-Depth: Monocular Depth Estimation Using High-Frequency Features to Enhance Structural Modeling

Abstract

Existing monocular depth estimation methods often struggle to preserve fine-grained spatial structures during feature encoding, which degrades the decoder’s ability to reconstruct accurate scene geometry. As a result, the predicted depth maps tend to exhibit structural blurring and depth discontinuities. To address this issue, we propose Hf-Depth, a high-frequency-enhanced monocular depth estimation network that improves structural representation by embedding high-frequency components into encoded features. Specifically, the dual-tree complex wavelet transform (DTCWT) is employed to extract high-frequency components from the input image, which are further refined through filtering and feature alignment. In addition, a channel selection attention fusion (CSAF) module is designed to effectively integrate encoded features with high-frequency features. Extensive experiments on public benchmark datasets, including KITTI, NYU Depth v2, and SUN RGB-D, demonstrate the effectiveness and competitive performance of Hf-Depth across different datasets. On the KITTI dataset, our method reaches a root mean square error (RMSE) of 2.163 and a $\delta _{{1}}$ accuracy of 0.972, achieving competitive performance compared with existing methods. Consistent improvements are also observed on the NYU Depth v2 dataset. The source code is currently private and will be made publicly available upon acceptance of this article.