Camouflage detection: experiments and a principled theory

Abstract

Camouflage is an impressive feat of biology in which an animal’s surface evolves to match the reflectance and texture of the backgrounds against which it typically appears. Equally impressive is the ability of visual systems to detect such camouflage. We present a principled theory of camouflage detection based on task-relevant cues and biologically plausible visual computations. This theory is informed by a series of psychophysical experiments where we measured human ability to detect maximally-camouflaged targets: ones whose texture is a random sample of the background texture. The amplitude spectra of natural images fall inversely with spatial frequency raised to an exponent, which varies from approximately 0.7 to 1.5 (this represents the degree of spatial correlation in the image). To see how this would influence camouflage detection, we measured detection on Gaussian noise textures with different falloff exponents. For 1o targets in the fovea, we find that humans are about 75% correct for an exponent of 0.7, and almost 100% correct for exponents above 1. We also find that performance degrades substantially in the periphery: e.g. at 12o eccentricity, humans only reach 75% correct when the exponent is 1.5. So interestingly, humans cannot detect maximally-camouflaged targets for exponents just below the range that occurs in nature. In other experiments, we measured camouflage detection on a variety of naturalistic textures, and also as a function of the complexity of the target shape. As a starting point, we focus our detection theory on only the information available at or near the target-background edge, so we exclude textures with strong long-range patterns that give away additional cues. We find that a principled model that includes edge-element detection at multiple scales, edge-element grouping, weak signal suppression and decision noise, can account for many aspects of our parametric experimental measurements.

Full text

•The models described aim to capture and measure all available information in
the stimuli.
•We are now building amodel that instead aims to match human detection
mechanisms:
1. Filter image with the human contrast sensitivity function
2. Detect all edge contours in the image
3. Separate them into boundary and texture contours
4. Measure several edge contour properties:length, edge
power at different sizes, curvature etc.
5. Discriminate target and blank stimuli in the space of all
these cues, allowing some noise in each cue
6. Fit this model to human detection data across all textures
= 0:
white noise
blank
target
•Most psychology/machine vision research uses
several cues to separate objects from background.
•Camouflage evolved to explicitly defeat these cues.
•So it exposes different detection strategies and
their limits.
•This moth camouflages maximally by matching all
background properties: brightness, contrast,
colour,texture.
•What is the available information here?
•How is it visually processed into adecision?
•When will detection be easy/hard?
•Our goal is to answer these, by:
•measuring human camouflage detection
against important stimulus parameters
•developing aprincipled detection model
Camouflage detection: experiments and a principled theory
Abhranil Das and Wilson Geisler
We design a stimulus & quantify its edge power
Detection on natural textures
Why study camouflage detection?
Key results
The edge power measure cleanly separates the target from the blank stimuli, and
sorts them by detectability.So we can do apsychophysical detection experiment
using these sorted stimuli, and measure the detection threshold:
Adding edges of different sizes
•We develop models to measure the main detection information in maximally-
camouflaged stimuli:the boundary edge.
•We experimentally measure human detection while varying important
parameters of the stimulus
•We show applications, such as calculating the best/worst hiding spots, or to
rate/compare different camouflage textures.
•‘Whitening’, i.e. flattening the stimulus amplitude spectrum, changes the
texture to information-less white noise, and losslessly pulls in the useful
information from all scales into a fine edge along the target boundary.
•We find that human detection is identical on the original pink noise stimuli
and their whitened versions.
•Human vision is known to detect and
combine edges of different sizes.
•We extend our model to compute edge
powers of stimuli at different scales,
and discriminate using all of them.
•We can optimally combine these edge
measures at different scales into a
single measure.
•This improves the detection
performance of our model.
decision
boundary
pink (1
) noise texture: same avg.
spectrum as natural images
edge
gradients
edge power: =2()
Experiments: detection as a function of edge power
blank
target
Adding the edge power spectrum
•The human visual system is known to detect localized clumps of gradients
that form continuous edge contours
•We can measure this using the power spectrum of the edge.When the
gradients clump together in groups, low frequencies will have more power.
•Using both the edge power and this power spectrum, detection improves:
frequency
edge power spectrum
target
blank
edge
gradients
()
location along boundary
threshold 
subject 1
subject 2
Detection of different shapes
•The shapes are filtered noise functions with different amplitude-spectrum
exponents, which determine their smoothness.
•We find that detection is hardest when the smoothness (exponent) of the
shape is similar to the smoothness (exponent) of the texture.
0
0.5
0.2
0.1
threshold
luminance
contrast
1
0.4
0.2
0.6
0.3
Detection threshold varies with stimulus parameters
Detection across filtered noise textures
0 200 400 600 800
stimulus duration (ms)
0.1
0.2
0.3
0.4
threshold
avg.
human
fixation
subject 1
subject 1
subject 2
•Beyond certain luminance/contrast/size/
duration, detection performance saturates.
•We find that detection reaches its best at
the average human fixation duration.
threshold
•The amplitude-spectrum exponent of natural
images varies from ~0.8 to ~1.2.
•We find that detection rises from just above
chance to nearly perfect in this same range.
Extending the model
Whitening pulls all information into a fine edge
00.05 0.1 0.15 0.2 0.25 0.3 0.35
edge power
50
75
100
% correct
1
noise stimuli
whitened stimuli
subject 1
Building a human detection model
paper (PhD thesis) video explanation
this poster
•Percentages denote the overall detection accuracy, using which the textures
are sorted from most to least detectable.
•Edge power does not fully determine detectability.Other factors matter too,
e.g. disruptive texture edges.