Monocular Depth
What a single-image depth network can know, what it cannot, and what the metrics quietly hide.
Geometry says depth is unrecoverable from one image. Monocular depth networks work anyway — by learning a prior over scenes, not by solving the geometry. That distinction is not philosophical: it determines exactly which parts of the output you can trust, which transfer to a new camera, and which are artefacts of the evaluation protocol.
This topic is the bridge between the projective geometry above and the depth maps, point clouds and metrics in the rest of this application.
- 01 What the network actually predicts Core 1 interactive 9′ Not depth. A depth-like quantity, defined up to a transformation nobody wrote down. Repairs Treating a network output as metres because the dataloader called it "depth" and the colourmap looked plausible.
- 02 Scale, shift, and the alignment protocol Core 1 interactive 11′ A two-parameter fit stands between the model output and every number in the results table. Repairs Comparing depth metrics across papers without knowing whether the alignment was scale-only, affine, in depth space, or in disparity space.
- 03 From depth map to 3D Core 1 interactive 9′ A depth map plus the wrong intrinsics produces a confident, wrong reconstruction. Repairs Blaming a bent or skewed point cloud on the depth model when the depth values are fine and the camera parameters are not.