Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Response to Affine Transforms of Image Distance Metrics and Humans

Paula Daudén-Oliver¹, Nuria Alabau-Bosque¹, Jorge Vila-Tomás¹, Jesus Malo², Valero Laparra³; ¹Universidad de Valencia, ²Universitat de Valencia, ³Universitat de València

Presenter: Paula Daudén-Oliver

The standard approach to testing image quality models with deep architectures is through correlation with human opinion of distortions typically found in digital media. RAID-database presents a more human way of testing distorted images with affine transformations that are the ones found in nature. We have selected 6 image quality metrics (2 convenient references and 4 state-of-the-art) to test their alignment with human behavior with the same psychophysical method: Maximum Likelihood Difference Scaling. Although perceptual metrics are designed to predict human perception, we found that none of them accurately replicate human response curves for the three proposed affine transformations. Specifically, we analyzed the ranking regard human responses to different images within the same distortion, the ranking with regard human sensitivity of single images when different affine distortions are applied, and the shape of the MLDS response curve.

Topic Area: Visual Processing & Computational Vision

Extended Abstract: Full Text PDF