A well-organized reading list can be more useful than one impressive demo. Awesome Deep Stereo Matching collects methods, datasets and related problems in one place. Looking through it is a reminder that stereo vision has plenty of unfinished work, even while single-image depth models receive much of the attention.
Two cameras provide a specific kind of evidence. A point appears at different horizontal positions in the left and right images. With calibrated, rectified cameras, that disparity can be related to distance. In the simplest pinhole model, depth is focal length times baseline divided by disparity.
The equation is compact. Finding the correct disparity is where most of the difficulty lives.
A blank wall has few distinctive features to match. A repeating fence offers too many similar candidates. A shiny surface can show different reflections to each eye. At an object's boundary, one camera may see background that the other camera cannot see at all. Those are different failures, and a single average error does not explain which one a method handles well.
Learn the ambiguity, keep the geometry
A learned stereo method can use more context than a small patch matcher. It can combine texture, edges and broader image structure to choose a plausible correspondence. That can produce cleaner results, but the geometry still has to be correct. A model should not be expected to repair an arbitrary camera alignment or a wrong baseline without being designed for that problem.
I find it useful to separate three questions when reading a result. Were the cameras and images prepared correctly? How well does the matching method recover disparity? What does the application do with uncertain or missing values?
For example, a depth map used to blur a background can tolerate some errors that a measurement tool cannot. A reconstructed scene may hide small noise on a wall but make an incorrect foreground edge conspicuous. A robot moving through space has yet another set of priorities. There is no single useful definition of "good depth" without a downstream task.
Time adds another dimension. A clean map in one frame can flicker badly in a video. If a surface jumps between two depths, a viewer sees instability even when the per-frame average looks acceptable. A practical evaluation should therefore include sequences, especially around occlusion boundaries and moving objects.
The reading list is a map, not a recommendation to choose the newest entry automatically. I would use it to locate methods with relevant assumptions, then inspect their original papers and code. Training data, camera resolution, runtime conditions and the treatment of invalid pixels all affect how a result transfers to another setup.
Stereo remains attractive because it connects appearance to a physical camera arrangement. When the calibration is trustworthy, the baseline gives the distance estimate a scale that a single image cannot supply on its own. That advantage is worth preserving carefully. A smarter matcher is useful; a clear account of what the cameras actually observed is still essential.