Answers are grounded in this paper and cite their sources.
Recovering geometrically coherent depth from underwater images is difficult because attenuation and scattering weaken local appearance cues and obscure object boundaries. We propose Neptune-Net, a dual-stream network that couples frequency-aware depth representations with segmentation-derived structural priors. A wavelet-based encoder–decoder retains low-frequency surface structure and directional detail across feature resolutions. A frozen EfficientSAM encoder supplies object-related context, which guides adaptive feature refinement through spatially varying residual modulation. This mechanism integrates structural information while preserving a depth-specific feature pathway, and an adaptive-bin head converts the refined representation into continuous depth predictions. Input wavelet enhancement and training-only WaterCut further address low contrast and local detail loss without altering depth supervision. Experiments on FLSea and USOD10K yield absolute relative errors of 0.1100 and 0.5150, respectively, with boundary F-scores of 0.1400 and 0.3800 at a two-pixel tolerance. On FLSea, Neptune-Net reduces absolute relative error by 28.57% relative to a controlled EfficientNet-B5/adaptive-bin baseline with the same input processing. Component ablations support the complementary contributions of wavelet processing, segmentation-conditioned refinement and input adaptation. These results demonstrate the effectiveness of coupling local frequency information with structural context for underwater monocular depth estimation.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: