Computational Approaches to Visual Illusion
Impossible Motions: A New Type of Visual Illusion Generated by Shape-from-Image Equations
A new type of visual illusion, which we call “impossible motion” is presented. In this illusion, we are given a solid object which looks like an ordinary shape, but motion added to the object will generate an impression that such motion cannot arise because they are against physical laws. Examples are balls rolling uphill along slopes defying the gravity law, and rods penetrating through two or more windows simultaneously defying straightness of the rods. These illusions are generated by utilizing the degrees of freedom in the choice of three-dimensional solid structures from two-dimensional images. For a given image of a familiar solid object, humans usually perceive one solid object, although there are infinitely many solid structures which yield the same image. Utilizing this gap between human perception and geometric constraints, we can cheat humans, thus designing “impossible motion”. We present many examples of impossible motion, and try to elucidate a nature of human perception of solid structures from two-dimensional images, in which some structures are preferred to others. In particular we present a hypothesis that human vision prefers highly symmetric structures, and show that the illusion of impossible motion can be explained by this hypothesis.
Shape-Free Hybrid Image - Effects of Artificial Noise and Complementary Color
Peeraya Sripian and Yasushi Yamaguchi
We present a new scheme for generating a shape-free hybrid image, an image that changes interpretation according to the viewing distance. A hybrid image is created by the combination of the low and the high spatial frequencies of two source images. It is based on human visual perception which perceives up to some range of spatial frequency at a specific visual angle. Our methods allow the construction of hybrid image regardless of the source image’s shape. Without the need to carefully pick the two images to be superimposed, a hybrid image can be extended to use with any kind of image contents. We propose two approaches to accomplish shape-free hybrid image, which are noise-inserted approach and color-inserted approach. Noise-inserted approach forces observers to perceive alternative low frequency image as meaningless noises in a close viewing distance, by manipulating contrast and details in the high frequency image and also by pre-process both source images before extracting spatial frequencies. Color-inserted approach attracts visual attention for the high frequency image perception by using complementary chromatic sinusoidal gratings. Finally, hybrid image recognition experiments prove that our proposed method yield a better recognition rate over the original method while preserving hybrid image characteristic.
Perceptual stabilization of ambiguous visual input: a synthesis of perception, computation and neurophysiology
Ambiguous visual stimuli contain sensory evidence for two (or more) mutually exclusive perceptual interpretations. While perceptual awareness is dominated by a single interpretation at any particular moment, dominance tends to alternate between different interpretations over prolonged viewing time. These dominance fluctuations can however be slowed down significantly by presenting ambiguous stimuli in sequences of brief presentation periods separated by interstimulus periods without visual input. The neural mechanisms that determine perceptual dominance at stimulus onset and the dynamics of perceptual alternations may help us understand the basic neuronal operations of perceptual organization. Here we present results from computational modeling, human psychophysical experiments, and neurophysiological recordings from monkey visual cortex, all aimed at understanding these mechanisms. The original computational work yielded the hypothesis that dynamics of perceptual dominance should crucially depend on the temporal profile of stimulus presentation. Behavioral experiments confirmed this hypothesis, refined the computational model and provided a handle for neurophysiological recordings. Data from these recordings revealed a range of effects on neuronal response variability that push the computational framework towards incorporating intra- and intercellular neuronal dynamics.
Illusions in Man and Machine
From a computational point of view, many of the processes involved in the interpretation of images are estimation processes and can be analyzed using the tools of statistics and signal processing. Through analyses of early visual computations we found intrinsic limitations in many processes, which make it impossible to compute veridical estimates in all imaging situations. Specifically, we found three principles governing the estimation of static image features and image motion. These are (a) statistical biases affecting the estimation of all image features, which can account for many geometric optic illusions and motion patterns such as the Ouchi illusion; (b) asymmetry in the filters computing temporal derivatives, which can account for illusory motion in patterns with asymmetric intensity signal such as the Snake illusion; (c) effects from compression of the signal, which can account for errors in the estimation of lightness and color illusions. Since these limitations are inherent to the computations, we argue, that they will affect artificial as well as biological systems. To understand these limitations, can help us improve our machine vision methods when they are designed for constrained environments.
Computational Creation of a New Illusionary Solid Sign with the Hollow Structure
We present a new illusionary solid sign, so-called "hollow arrow sign", inspired by two kinds of illusions; "hollow mask illusion" and "crater illusion". This illusionary sign creates a visual illusion in such a way that the depth of the solid is inversely perceived for one's eyes due to the illumination direction. Moreover, the hollow sign appears to move in the same direction as the observer when he / she changes his / her observation point. In general, the productions to provide the visual illusion are obtained on an empirical basis. However, anyone can create this kind of the illusionary solid sign that is a solid sign with the hollow structure by our computational method to obtain the three-dimensional vertices of the illusionary solid.