Reproduce “Understanding SSIM” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jim Nilsson, Tomas Akenine-Möller
A paper by Jim Nilsson and Tomas Akenine-Möller from NVIDIA critically scrutinizes the Structural Similarity Index (SSIM), uncovering its mathematical limitations and frequent mismatches with human visual perception. The analysis demonstrates SSIM's true theoretical range, the potential for undefined results (NaNs) in its variants, and its counter-intuitive behavior across various image distortions.
Key insights
- SSIM's theoretical range is (-1, 1], not [0, 1], and its 's' (structure) component can produce negative values which, when raised to non-integer powers (e.g., in MS-SSIM), yield complex numbers or NaNs.
- The luminance component of SSIM exhibits non-intuitive perceptual behavior, being overly sensitive to small changes near black and conversely insensitive to noticeable differences near white, contradicting human perception.
- SSIM frequently fails to accurately reflect perceived image quality, ignoring significant chrominance shifts or structural distortions while sometimes penalizing visually subtle differences disproportionately.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.