Many new AI models estimate how far away things are in metres from a single photo, a task called metric depth estimation. To do this, they must also figure out the camera that took the picture, for example its focal length. A simple test reveals a flaw: when a photo is digitally zoomed, which is just cropping and enlarging, the scene does not actually move, so the distances reported for every point still in the picture should stay the same.
The researchers tested this by cropping 800 photos from standard datasets and enlarging the crops back to full size. They measured two things for each model: how depth changed with zoom, and how the model's estimate of focal length changed. The ideal model would have depth unchanged and focal length scaling perfectly with zoom.
The widely used model Depth Anything V2 showed depth changes almost as if the camera had moved closer, with a value of 0.95 out of 1 for the depth-change measure. A simple fix using the true camera made the error vanish, showing the root cause is wrong camera estimates. This means the models learn a shortcut from object size to distance rather than understanding geometry.
