Depth Estimation from Hyperspectral Images Based on Chromatic Aberration

Chromatic aberration results in the focal length varying across different wavelengths, leading each wavelength to focus on objects at different distances. Based on this phenomenon, we propose a depth estimation network based on spectral chromatic aberration. Since there are no publicly available datasets for training spectral depth estimation networks, we constructed a large-aperture spectral imaging system without chromatic aberration correction to acquire hyperspectral images. The corresponding ground-truth depth maps were obtained using monocular structured light 3D measurement technology. The proposed spectral chromatic aberration-based depth estimation network is fundamentally based on an encoder-decoder architecture. To better extract features from the spectral cube and enable deeper encoding, we introduce CDC Blocks and LGFI Block during the downsampling stages, and construct the encoder in five stages. To improve depth prediction accuracy, intermediate depth maps generated at each decoder stage are concatenated with corresponding encoder features via skip connections. Experimental results on the collected dataset demonstrate the effectiveness of the proposed method, with relative errors below 10% and accuracy exceeding 85%.