Monocular Depth Estimation From the Perspective of Feature Restoration: A Diffusion Enhanced Depth Restoration Approach

A research team led by Huibin Bai and Shuai Li has published a novel paper proposing a significant shift in how AI models perform Monocular Depth Estimation (MDE)—the task of predicting a 3D depth map from a single 2D image. The core innovation is reframing the problem from direct prediction to one of feature restoration. The researchers argue that features extracted by a standard encoder are effectively degraded versions of an ideal 'ground truth' feature. To restore these features, they developed the Invertible Transform-enhanced Indirect Diffusion (InvT-IndDiffusion) module, a specialized diffusion model that operates under a bi-Lipschitz condition to maintain stability during its iterative process, solving a key issue of feature deviation.

This main module is complemented by a plug-and-play Auxiliary Viewpoint-based Low-level Feature Enhancement (AV-LFE) module, which uses available multi-view data to sharpen local details. The combined system was tested on standard benchmarks, with standout results on the challenging KITTI dataset for autonomous driving. The paper reports a remarkable 37.77% improvement in the key RMSE (Root Mean Square Error) metric under one training setting, and a 4.09% gain in another, compared to their baseline. This indicates there is substantial, untapped performance potential within the common encoder-decoder architecture used by most current MDE methods, simply by improving how encoder features are processed and refined.