Camera-independent depth from defocus
Single-image depth estimation from defocus blur that keeps working when you change the camera. WACV 2024.
Monocular depth-from-defocus models learn how blur relates to depth, but that relationship depends on the camera (focal length, aperture, sensor). A model trained on one camera quietly breaks on another.
Using the optics of defocus blur, I showed how camera parameters change the blur, and built a correction that lets a trained model generalize to new cameras without retraining. I also built a synthetic dataset to test camera independence, and evaluated on it as well as DDFF12 and NYU Depth V2 captured with different cameras.