Reproduce “Monocular absolute depth estimation from endoscopy via domain-invariant feature learning and latent consistency” on OpenResearch

About this paper

Hao Li, Daiwei Lu, Jesse d'Almeida, Dilara Isik, Ehsan Khodapanah Aghdam, Nick DiSanto, Ayberk Acar, Susheela Sharma and 3 more

Researchers at Vanderbilt University developed a latent space feature alignment method for monocular absolute depth estimation in endoscopy. This technique, which combines domain-invariant feature learning and latent consistency, reduces the domain gap between synthetic and real endoscopic images, leading to improved depth accuracy for guiding autonomous medical robots.

Key insights

• Jointly training a shared depth network with combined supervised (translated synthetic) and unsupervised (real) data, coupled with feature alignment, effectively bridges the residual domain gap more robustly than prior methods.
• Domain-Invariant Feature Learning (DIFL) and Latent Consistency (CON) work synergistically: DIFL aligns global feature distributions, while CON refines this alignment by ensuring semantic correspondence.
• The proposed method allows for improved performance and robustness across various backbone architectures, indicating that aligning features directly enables better utilization of model capacity without overfitting to domain-specific artifacts.

Summary by alphaXiv.