Abstract
Camouflage must be appreciated as an extraordinary feat of evolution, but any detection of such camouflaged objects by predator and prey is also an impressive feat of visual systems. Many historical studies of camouflage have focused on descriptive compilations or heuristic applications. Our goal instead is to understand camouflage detection in the framework of controlled psychophysical experiments, and to develop a principled theory based on task-relevant stimulus statistics and known biological vision mechanisms. Moreover, unlike most object detection questions in computer science and psychology that utilize multiple cues, we focus on the particularly hard scenario where the camouflaging object exactly mimics the luminance, contrast, color and texture of its background. What then is the available information, and how does the visual system exploit it?
We recognize that most of the information here resides at the object-background edge. We define measures to quantify the total magnitude and spatial distribution of this edge information, calculate them at multiple spatial scales in accordance with known early visual computations, and develop a method to condense these correlated cues into fewer dimensions. We also describe a whitening procedure that decorrelates the texture and losslessly gathers the distributed cues into a narrow object boundary.
In parallel, we characterize human psychophysical detection performance on stimuli with pink noise texture (which is well-studied and shares properties of natural scenes, hence provides a principled starting point), then extend to more naturalistic textures. We find that the edge measure that we have developed predicts detection performance smoothly, allowing us to extract detection thresholds across differing conditions of luminance, contrast, target size and stimulus duration.
We apply these findings to identify the best location for an object to hide against a background, and evaluate the effectiveness of different textures for camouflage.