Alexander Hepburn
Postdoc at University of Bristol
Perceptual distances model a distance between two images, and are often designed to replicate certain processes in the human visual system, or optimised to mimic the decisions in a set of human perceptual judgements. Assuming Barlows hypothesis, that the brain seeks to minimise the amount of redundant information in a signal, the outcome of these perceptual judgements should be related to the probability of the images.
We investigate this claim through the application of Riemannian distances in the context of image perception modeling. Leveraging the probability of images, we extend the image space to include the image probability and define Riemannian distances as the shortest geodesic curves connecting images within this augmented domain. To overcome the computational challenges in high-dimensional spaces, we will introduce a discrete-based piecewise-linear approximation. This approach optimises point placement along the trajectory between two images to minimise the Riemannian distance. We also explore local approximations given the proximity of images, where we have image pairs consisting of a reference and a distorted version. We evaluate our distance models by predicting human perceptual judgments. Notably, our proposed distance model offers interpretability, enabling visual inspection of points along the curve to explain decision-making processes, rather than relying on complex transformations that are difficult to visualise.
We also propose a novel method to evaluate perceptual distances from a set of perceptual judgements from the two-alternative forced choice experiment, by assuming a binomial distribution models the decision process.
The visit is supervised by Javier Portilla at the Instituto De Optica – CSIC.
Keywords: human perception, natural images, Riemannian manifold, geodesic distance, perceptual experiments
Scientific area: Image Perception, Artificial Intelligence
Bio: Alex Hepburn is a senior research associate at University of Bristol in the Machine Learning and Computer Vision Research Group, focusing on human visual perception in machine learning. Alex completed his PhD at University of Bristol in 2022 on including perceptual information in the loss functions, architecture and explanations of machine learning systems.
Visiting period: January 2024 at Instituto de Optica CSIC