Advancing Monocular Depth Estimation: High-Resolution and Amodal Understanding

This thesis advances monocular depth estimation along two complementary directions. It develops efficient methods for recovering fine geometric details, such as object boundaries and thin structures, from high-resolution images while preserving scene-wide consistency and metric scale. It also extends depth estimation to partially occluded objects by predicting the relative depth of their surfaces across the full region they project onto the image, including parts hidden from view.

A single RGB image contains cues about the 3D structure of a scene, from the overall layout of surfaces to the fine contours of individual objects. Modern depth models can recover much of that layout, but they typically downsample high-resolution inputs to fit practical memory and compute budgets. Downsampling can erase boundaries and thin structures before the model sees them, and upsampling the resulting depth map cannot reliably bring them back. Learning these details from real-world data is also difficult because depth annotations are often least reliable around the structures that require the most precise supervision.

The thesis first addresses this problem with PatchFusion, which combines the scene-wide consistency of a whole-image depth prediction with the detail recovered from local high-resolution patches. The framework learns to fuse the two sources of information so that neighboring patches agree in scale and structure. PatchRefiner builds on this approach by learning to refine a coarse depth prediction and by using a teacher trained on synthetic data to generate detailed depth pseudo-labels for real images. A training objective separates local detail recovery from metric scale, allowing the model to sharpen boundaries while preserving reliable depth values. PatchRefiner V2 makes the approach more practical by replacing the second full-sized depth model with a lightweight refinement branch that uses coarse scene information to guide the recovery of fine details.

The final part of the thesis asks what depth estimation can recover when part of an object is hidden. Amodal Depth Anything estimates the relative depth of a target object across its full projected image region, including the pixels where another object blocks it from view. The prediction stays aligned to the input image, so it describes where the hidden surface lies in depth at those pixels. To study this task, the thesis introduces a large-scale training dataset for everyday scenes and develops both deterministic and generative models for predicting occluded geometry.

Together, these works improve monocular depth estimation in two ways. They preserve finer geometry where the image provides evidence and infer the depth of object surfaces where that evidence is incomplete.