Physics-informed deep optics: chromatic aberrations for monocular depth estimation

Absolute monocular depth estimation remains fundamentally ill-posed due to the loss of metric scale. While deep optics offers a physical resolution by co-designing optics and algorithms, existing end-to-end frameworks often lack physical interpretability, rely on unrealistic shift-invariant point spread function (PSF) approximations, and yield difficult-to-manufacture freeform elements. To address these limitations, we propose a physics-informed, proxy-driven deep optics framework that repurposes chromatic aberration from an imaging defect into a robust metric-depth cue. To mitigate the unconstrained nature of end-to-end learning, we optimize manufacturable spherical lenses via differentiable ray tracing guided by explicit proxy losses, actively distributing wavelength-dependent focal responses across the diopter space. Furthermore, we eliminate the shift-invariant assumption by introducing a dual soft-interpolation rendering model to generate physically rigorous, spatially variant degraded images. These encoded cross-channel chromatic gradients are then explicitly decoded by a lightweight, optical-aware U-Net. Evaluated on the NYU Depth v2 benchmark, our 6-element spherical lens configuration (F/4 aperture) achieves an RMSE of 0.1821 m. Notably, by delegating the complex task of absolute scale estimation to the optical front-end, our physically constrained system empowers a mere 25M-parameter decoder to achieve absolute metric accuracy that rivals massive vision foundation models, establishing a highly efficient and manufacturable paradigm for passive 3D sensing.