Mapping Parameter Space Using Human Detection Metrics

Abstract

Camouflage development traditionally relies on comparing a small number of handpicked patterns in human detection experiments or, more recently, on automated evaluations using computer vision models. Both approaches come with their own limitations: the former depends on a restricted number of candidate patterns which may suffer from a researcher’s bias, while the latter may fail to reflect human perception. These methodological constraints hinder systematic exploration of how camouflage pattern parameters interact to influence detectability by human observers.

This thesis introduces an alternative method that captures human detection performance via a sampling process across a continuous N-dimensional parameter space, where each parameter combination defines a unique camouflage pattern (in this study N = 4). The corresponding human response is then modelled as a noisy observation from an underlying latent detection-difficulty function, allowing the modelling of how individual parameters and parameter interactions shape overall detectability.

Blender’s render software and Python API integration were used to generate a large volume of fully synthetic, parametrically defined stimulus samples (18000 unique images in total). Since fully synthetic camouflage visualizations can be generated with precise, granular control and at negligible cost, this approach enables dense, non-repeating, human-in-the-loop sampling across multidimensional parameter variations. This study shows evidence that a machine learning model can map the parameter space and predict camouflage performance. We also provide recommendations which should enable a significant improvement in model accuracy and reduce the number of trials needed to reach saturation, as well as suggestions for other areas where this method and tool can be applied