Guessing depth from a single frame
Monocular depth estimation trains a deep learning model to predict a depth value for every pixel in a single ordinary image, no second camera and no dedicated sensor required. This sounds like it should be impossible given everything covered in the first lesson: a single photo genuinely does not contain enough information to derive depth through geometry alone. What makes it work is that the model is not doing geometry, it is doing pattern recognition, learning statistical regularities about how the visual world tends to be arranged from being trained on huge datasets of images paired with real depth measurements, often collected using the very LiDAR and stereo systems covered in the previous lessons.
These are largely the same cues people use without realizing it. Objects positioned higher up in the frame and rendered smaller are usually farther away, this is why a monocular model can make a reasonable depth guess about a road stretching to the horizon. Surface textures, gravel, grass, brick, appear less detailed and finer-grained as distance increases, giving the model another signal. And occlusion, one object partially covering another, unambiguously tells you the covering object is closer, since a farther object cannot physically block a nearer one.