level: research
learning-based single-shot fringe projection profilometry usually works at close range. at long range, beyond one meter, the problem gets harder. light intensity drops with distance, lowering signal quality. a single image lacks fringe-order information, making the task ill-posed. researchers ran a diagnose-repair-verify study on a synthetic benchmark with 15,600 fringe images of 50 objects at 1.5 to 2.1 meters. a baseline unet model gave 14.54 mm mean absolute error.
they used mechanistic interpretability and conformal uncertainty quantification as diagnostics. three probes—linear probing, grad-cam, and a flat-plane out-of-distribution test—all pointed to the same failure. the network relied on a shape-prior shortcut instead of learning true depth from fringes. this shortcut worked on training-like shapes but failed on new geometries. the agreement between interpretability and uncertainty methods confirmed the physical cause of the error.
the repair involved architectural changes to remove the shortcut. after fixing, the model's error dropped significantly. the study shows how combining interpretability and uncertainty tools can find and fix hidden failures in learning-based 3d imaging. this approach could apply to other ill-posed vision problems where networks learn spurious correlations. the work advances reliable single-shot profilometry for long-range applications like industrial inspection or autonomous systems.
why it matters: it shows how to make neural networks for 3d imaging more reliable by finding and fixing hidden shortcuts, which is crucial for safety-critical ai applications.