Monocular Depth Estimation

Monocular depth estimation predicts how far visible surfaces are from a camera using visual cues such as perspective, texture, shading, and learned scene structure. Methods range from supervised learning with depth maps and self-supervised training from video or camera motion to foundation models, diffusion models, and geometric or physical cues. Recent work adapts these models to different cameras and domains, adds sparse sensor measurements to recover metric scale, and extends prediction to video, panoramic images, and specialized settings such as surgery or aerial imaging. A central challenge is making predictions both metrically accurate and temporally consistent while retaining detail and generalizing across cameras, environments, and limited-compute devices.