Estimating the 3D structure of a scene from a single 2D image—monocular depth estimation—is a fundamental challenge in computer vision. While humans use a variety of cues like perspective, occlusion, and known object sizes to perceive depth, training a neural network to do the same typically requires massive amounts of ground-truth data, such as LiDAR scans, which are expensive and difficult to collect.
The research direction explored across these papers shifts the paradigm from supervised learning to self-supervised geometric consistency. Instead of telling a network what the depth is, researchers force the network to predict a depth map that makes multiple views of the same scene "mathematically consistent."
The Core Intuition: View Synthesis as Supervision
The fundamental trick behind this field is treating depth estimation as an intermediate step in an image reconstruction task. If you have two images of the same scene taken from slightly different positions (a stereo pair), and you know the distance between the cameras, the only missing piece of information to transform the left image into the right image is the depth of every pixel.
By training a network to predict a depth map that allows us to "warp" one view to match another, we create a supervisory signal without ever needing a LiDAR sensor. Yevhen Kuznietsov and colleagues demonstrated that this unsupervised stereo alignment can be combined with sparse ground-truth data to produce highly detailed depth maps that exceed the resolution of the original sensors.
Distilling Knowledge from Stereo to Monocular
A recurring theme in this research is the use of Knowledge Distillation. Stereo matching is a much easier geometric problem than monocular depth estimation because the "search" for a pixel's match is constrained to a single line.
Researchers like Hongsheng Li and Jimmy Ren leverage this by using a sophisticated "Teacher" stereo network to guide a "Student" monocular network. The Teacher looks at two images and generates a proxy depth map, which the Student then tries to replicate using only a single image. Andrea Pilzer refined this further by introducing "cycle-inconsistency," where the model identifies areas where the reconstruction fails and uses those errors to self-correct during the distillation process.
Bridging the Gap with Depth Hints and Priors
Self-supervised models often struggle with thin structures or low-texture areas where the photometric (visual) match is ambiguous. To solve this, Jamie Watson and Gabriel Brostow introduced Depth Hints. Instead of relying solely on the neural network's predictions, they use "noisy" suggestions from classical, non-learned stereo algorithms. The network is trained to follow these hints only when they result in a better image reconstruction, effectively using 50 years of traditional computer vision to guide modern deep learning.
Similarly, Fabio Tosi and Matteo Poggi developed monoResMatch, which forces a monocular network to "pretend" it is a stereo system by synthesizing a virtual second viewpoint, allowing it to apply traditional stereo matching logic to a single-camera input.
Domain Adaptation and the Real-World Bottleneck
A major hurdle in this field is the "domain gap"—models trained on synthetic data (like video games) often fail in the real world. Alessio Tonioni and Stefano Mattoccia addressed this by using unsupervised domain adaptation, where a model fine-tunes itself on real-world stereo pairs by generating its own "noisy" labels. Shanshan Zhao expanded this with GASDA, which uses style transfer to make synthetic images look real while enforcing geometric constraints.
Despite these successes, a persistent bottleneck remains: the training requirements. While these models only need one image at test time, almost all current self-supervised methods—including those for panoptic segmentation by Christian Rupprecht or 6D pose estimation by Fabian Manhardt—still require synchronized stereo pairs or video sequences during training. This dependency prevents these models from learning from the billions of independent, static images available on the internet, a frontier that researchers like Ignacio Rocco are beginning to explore through multiview registration.