The article discusses the use of two-alternative forced choice (2AFC) experiments to evaluate perceptual distances. It compares traditional methods with crowd-sourced datasets like BAPPS, highlighting the challenges of ranking algorithms and the need for statistical modeling. The proposed method involves fitting a binomial distribution to perceptual judgments, allowing for consistency and smoothness in estimates. By utilizing maximum likelihood estimation, the study aims to assess different distances and calculate likelihoods of judgments based on empirical data. The experiments conducted on various candidate distances show promising results, especially with deep learning-based metrics like PIM, LPIPS, and DISTS. The robustness of the method is tested through hyperparameter tuning, demonstrating stability across different settings.
toiselle kielelle
lähdeaineistosta
arxiv.org
Tärkeimmät oivallukset
by Alexander He... klo arxiv.org 03-18-2024
https://arxiv.org/pdf/2403.10390.pdfSyvällisempiä Kysymyksiä