Emerging Images
Niloy J. Mitra, Hung-Kuo Chu, Tong-Yee Lee, Lior Wolf, Hezy Yeshurun, Daniel Cohen-Or


Emergence refers to the unique human ability to aggregate information from seemingly meaningless pieces, and to perceive a whole that is meaningful. This special skill of humans can constitute an effective scheme to tell humans and machines apart. This paper presents a synthesis technique to generate images of 3D objects that are detectable by humans, but difficult for an automatic algorithm to recognize. The technique allows generating an infinite number of images with emerging figures. Our algorithm is designed so that locally the synthesized images divulge little useful information or cues to assist any segmentation or recognition procedure. Therefore, as we demonstrate, computer vision algorithms are incapable of effectively processing such images. However, when a human observer is presented with an emergence image, synthesized using an object she is familiar with, the figure emerges when observed as a whole. We can control the difficulty level of perceiving the emergence effect through a limited set of parameters. A procedure that synthesizes emergence images can be an effective tool for exploring and understanding the factors affecting computer vision techniques.


(Top) This image, when stared at for a while, can reveal four instances of a familiar figure. Two of the figures are easier to detect than the others. Locally there is little meaningful information, and we perceive the figures only when observing the whole figures.
(Left) A classic example of an emergence image. Although at first sight the left image looks meaningless, suddenly we perceive the central object as the Dalmatian dog pops out.
(Right) Emergence images, when observed through small windows, look meaningless. Although we perceive the subject in the whole image, the smaller sized segments, in isolation, look like random patches. In contrast, the elephant can be recognized through similar windows of the normal shaded scene.
(Left) We often fail to perceive an emergence image when the subject is in an uncommon pose. Among the users who were shown the above images, the average success rate was only 54% and 4%, respectively. When the inverted versions of these images were shown, the success rates went up to 96% and 91%, respectively.

Mouse over the image to flip it.

(Below) Typical emergence images generated by our synthesis algorithm. We generate a range of examples on various subjects synthesized at different difficulty levels. Each example contains exactly one subject. (Please refer to supplementary material for other examples.)
(Beow) In many computer vision recognition or segmentation algorithms, the first stages comprise of multi-scale edge detection or other means of bottom-up region processing. At multiple-scales, we detect edges using standard Canny edge detector, and retain the ones that persists scales. Such curves are then linked together based on spatial proximity and curvature continuity. We observe that while on the original renderings the method successfully extracts the feature curves (right image in each box), on the emerging images the results can mostly be seen as noise. This indicates the difficulty that bottom-up algorithms face when detecting objects in the emergence images.
(Left) Emerging frog at various difficulty levels, increasing from left to right. We control the difficultly by controlling the sampling density, breaking the silhouette continuity, perturbing silhouette patches, and adding clutter using cut-perturb-paste.
(Right) Difficulty level as perceived by users and as predicted by our synthesis parameters. (Right) Perceived difficulty level in each category changes gradually. For example, 98% of the easy images were recognized by at least 80% of the observers.


AUTHOR = "Niloy J. Mitra and Hung-Kuo Chu and Tong-Yee Lee and Lior Wolf and Hezy Yeshurun and Daniel Cohen-Or",
TITLE = "Emerging Images",
JOURNAL = "ACM Transactions on Graphics",
VOLUME = "28",
NUMBER = "5",
YEAR = "2009",
NOTE = "to appear",

In News:

user study images
training dataset
demo app.
(win 5MB)
back to publications
back to homepage