Scale-Invariant Monocular Depth Estimation Via SSI Depth

Existing methods for scale-invariant monocular depth estimation (SI MDE)often struggle due to the complexity of the task, and limited and non-diversedatasets, hindering generalizability in real-world scenarios. This is whileshift-and-scale-invariant (SSI) depth estimation, simplifying the task andenabling training with abundant stereo datasets achieves high performance. Wepresent a novel approach that leverages SSI inputs to enhance SI depthestimation, streamlining the network's role and facilitating in-the-wildgeneralization for SI depth estimation while only using a synthetic dataset fortraining. Emphasizing the generation of high-resolution details, we introduce anovel sparse ordinal loss that substantially improves detail generation in SSIMDE, addressing critical limitations in existing approaches. Throughin-the-wild qualitative examples and zero-shot evaluation we substantiate thepractical utility of our approach in computational photography applications,showcasing its ability to generate highly detailed SI depth maps and achievegeneralization in diverse scenarios.