Coevolution of Camouflage

TL;DR

The paper presents an abstract 2D coevolutionary framework that models camouflage evolution as an adversarial dynamic between prey texture evolution and predator vision learning, guided by real background photographs. Prey textures are generated via TexSyn-based genetic programming, while predators are CNN-based detectors that learn from tournament outcomes and adapt over time; the competition uses a tournament-based relative fitness mechanism and periodic auto-curation. The study demonstrates that camouflage can emerge and sharpen under coevolution, provides an open-source testbed for exploring camouflage metrics and background-stimuli interactions, and suggests avenues for extending into 3D and alternative evaluation methods.

Abstract

Camouflage in nature seems to arise from competition between predator and prey. To survive, predators must find prey, and prey must avoid being found. This work simulates an abstract model of that adversarial relationship. It looks at crypsis through evolving prey camouflage patterns (as color textures) in competition with evolving predator vision. During their "lifetime" predators learn to better locate camouflaged prey. The environment for this 2D simulation is provided by a set of photographs, typically of natural scenes. This model is based on two evolving populations, one of prey and another of predators. Mutual conflict between these populations can produce both effective prey camouflage and predators skilled at "breaking" camouflage. The result is an open source artificial life model to help study camouflage in nature, and the perceptual phenomenon of camouflage more generally.

Figures (21)

• Figure 1: Photographs of natural textures, each overlaid with three camouflaged prey. The prey are randomly placed 2D disks, each with its own evolved camouflage texture. (Background photos of: plum leaf litter, tree and sky, gravel, oxalis sprouts. Zoom in for detail. Disk diameter is 20% of image width.
• Figure 2: Prey camouflage evolving over simulation time to become more effective in a given environment.
• Figure 3: Overview of one step of the coevolutionary simulation of camouflage. Three prey are selected at random from their population of 400. Similarly for three predators from their population of 40. A random background image is selected from the given set, and a random crop of 5122 pixels is made. The three prey are rendered over the background at random non-overlapping locations. This composite tournament image is given to each predator which estimates a position (circled crosshairs in tournament image, see also Figure \ref{['fig:predator_responses']}) predicting the center point of the most conspicuous prey. The predators are scored by "aim error" --- the distance from their estimate to the ground truth center of the nearest prey. If the best predator's estimate is inside a prey's disk, that prey is eaten and replaced by a new offspring of the other two prey. If all predators fail, all prey survive. If the worst scoring predator's estimate is outside all prey, it may die of starvation, to be replaced by a new offspring predator.
• Figure 4: TexSyn expression trees and crossover between them, illustrated here with a simplified version of TexSyn with just three texture operators (spots, stripes, and warp) plus four named solid color textures. Minimal operator trees are shown in (a) and (b). Crossover between (a) and (b) is shown in (c) and (d). (c) is spots where blue is replaced with stripes. (d) is stripes where gray is replaced with spots. (e) and (f) show (b) and (d) under a warp operator. (See Section \ref{['sec:cpp_code']} for the actual TexSyn c++ code used to create these examples.)
• Figure 5: Tournament image after simulation step: three camouflaged prey on a random background crop. Three crosshair marks show the responses of three predators, ranked by minimum distance to a prey center. Details in Section \ref{['sec:additional_predator_responses']}.