Rate-distortion optimization for full-reference image quality metrics via stochastic Hessian estimates
The paper reports 14.2-36.7% BD-rate savings when video encoding optimizes for perceptual image-quality metrics instead of SSE.
The authors target full-reference metrics such as MS-SSIM and LPIPS, which usually cannot be used directly inside block-based codec decisions. They approximate those metrics with an input-dependent quadratic distortion built from Hessian information at the source video. To make that usable block by block, they test block-diagonal and diagonal Hessian approximations estimated through automatic differentiation. In VVC tests on Kodak and CLIC, the method needs no decoder changes and adds 10-30% encoding complexity overhead. HF Daily Papers' note
The authors target full-reference metrics such as MS-SSIM and LPIPS, which usually cannot be used directly inside block-based codec decisions. They approximate those metrics with an input-dependent quadratic distortion built from Hessian information at the source video. To make that usable block by block, they test block-diagonal and diagonal Hessian approximations estimated through automatic differentiation. In VVC tests on Kodak and CLIC, the method needs no decoder changes and adds 10-30% encoding complexity overhead. HF Daily Papers' note
score 4