ICLR2025

Controllable Blur Data Augmentation Using 3D-Aware Motion Estimation

Insoo Kim, Hana Lee, Hyong-Euk Lee, Jinwoo Shin

摘要

Kernel-Based Methods (+) Estimate blur kernels to synthesize blur images (-) Simply regressing the simulated kernels may not hold in generating realistic blur images (-) The kernel-based methods are designed to estimate motions from a 2D perspective We may utilize byproduct of the kernel-based methods, i.e., blur kernels, to synthesize blur images J-MKPD [1] MotionETR [2]