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.

We introduce an example-based synthesis technique that extrapolates novel styles for a given input image. The technique is based on separating the style and content of image fragments. Given an image with a new style and content, it is first adaptively partitioned into fragments. Stitching together novel fragments produces a coherent image in a new style for a given content. The aggregate of synthesized fragments approximates a globally non-linear model with a set of locally linear models. We show the result of our method for various artistic, sketch, and texture filters and painterly styles applied to different image content classes.

Shape-from-shading (SFS) is a fundamental problem in computer vision. The vast majority of research in this field have assumed orthography as its projection model. This paper reexamines the basis of SFS, the image irradiance equation, under an assumption of perspective projection. The paper also shows that the perspective image irradiance equation depends merely on the natural logarithm of the depth function (and not on the depth function itself), and as such it is invariant to scale changes of the depth function. We then suggest a simple reconstruction algorithm based on the perspective formula, and compare it to existing orthographic SFS algorithms. This simple algorithm obtained lower error rates than legacy SFS algorithms, and equated with and sometimes surpassed state-of-the-art algorithms. These findings lend support to the assumption that transition to a more realistic set of assumptions improves reconstruction significantly.

We present a new method for completing missing parts caused by the removal of foreground or background elements from an image. Our goal is to synthesize a complete, visually plausible and coherent image. The visible parts of the image serve as a training set to infer the unknown parts. Our method iteratively approximates the unknown regions and composites adaptive image fragments into the image. Values of an inverse matte are used to compute a confidence map and a level set that direct an incremental traversal within the unknown area from high to low confidence. In each step, guided by a fast smooth approximation, an image fragment is selected from the most similar and frequent examples. As the selected fragments are composited, their likelihood increases along with the mean confidence of the image, until reaching a complete image. We demonstrate our method by completion of photographs and paintings.

Detection of feature points in images is an important preprocessing stage for many algorithms in computer vision. We address the problem of detection of feature points in video sequences of 3D scenes, which could be mainly used for obtaining scene correspondence. The main feature we use is the zero crossing of the intensity gradient argument. We analytically show that this local feature corresponds to specific constraints on the local 3D geometry of the scene, thus ensuring that the detected points are based on real 3D features. We present a robust algorithm that tracks the detected points along a video sequence, and suggest some criteria for quantitative evaluation of such algorithms. These criteria serve in a comparison of the suggested operator with two other feature trackers. The suggested criteria are generic and could serve other researchers as well for performance evaluation of stable point detectors.

This paper presents biological evidence for camouflage breaking using the convexity of the intensity function. Some animals use apatetic coloring especially to prevent their detection by gray-level convexity. This implies that other animals might be able to break camouflage based on gray-level convexity. We demonstrate the effectiveness of convexity-based camouflage breaking using an operator ("D/sub arg/") for detection of 3D convex or concave gray levels. Its high robustness and the biological motivation make D/sub arg/ particularly suitable for camouflage breaking. As will be demonstrated, the operator is able to break very strong camouflage, which might delude even human viewers. Being nonedge-based, the performance of the operator is juxtaposed with that of a representative edge-based operator in the task of camouflage breaking. Better performance is achieved by D/sub arg/ for animal as well as military camouflage breaking.

There is a recent growing interest in image analysis of multiple views of a scene, often involving aspects of reconstruction, mosaicing and new view generation. As the availability of multiple camera systems augments, we suggest that such tasks could be carried out where the image source is a set of unsynchronized and uncalibrated cameras moving arbitrarily in a 3D scene. In order to make efficient use of this data, it is necessary to define a measure of inter-sequence proximity. We suggest such a measure, based on pure 2D analysis, namely the ratios between image-space distances among a set of feature points. We show this measure to be sound, and propose a simple iterative method to robustly estimate the relative positions of the set of moving cameras, even in the presence of a substantial amount of noise, and without computing egomotion.

Computer vision systems are becoming more and more anthropomorphic. In addition to taking an active vision approach, researchers are now investigating robot heads with biologically motivated space-variant sensors. We have designed such an exploring eye system for robot gaze control when observing an arbitrary scene. Our approach uses "retinal" images constructed from overlapping receptive fields that are mapped to form log-polar "cortical" images. We have implemented this system on a parallel SIMD machine using an efficient parallelization algorithm. In order to demonstrate our approach, we have formulated various types of attentional operators that are used to select interest points to drive the gaze control mechanism. We show that our system efficiently integrates images from space-variant snapshots in an arbitrary natural environment.

Face recognition schemes that are applied directly to gray level pixel images are presented. Two methods for reducing the overfitting-a common problem in high dimensional classification schemes-are presented and the superiority of their combination is demonstrated. The classification scheme is preceded by preprocessing devoted to reducing the viewpoint and scale variability in the data.

Presents a procedure for facial normalization which is determined by the translocation of few anchors in two sets. The first set determines the affine part of the transformation and is used for overcoming change in viewpoint and scale. The second set is used for normalization of expression. A generalization of radial basis functions theory is used for generating the appropriate mapping. Only a few (1-3) anchors were used to demonstrate elaborate facial expressions without any face model.

Texture consists of local variance of gray level or edge intensity values. The authors have recently presented a generalized symmetry operator that captures local spatial relations of image patterns. They show that activity differences in the continuous intensity map produced by the local generalized symmetry operator can be efficiently used to detect texture boundaries. Using almost all available quantitative results of human performance in artificial texture discrimination, the authors show that the algorithm favorably compares with other computational approaches. Stressing the necessity of benchmarks for computer vision algorithms, the authors also discuss an exhaustive set of textures that could be used as experimental stimuli for both humans and machines. The performance of the algorithm is demonstrated on some of these artificial textures as well as on natural images.>

Locating facial features is crucial for various face recognition schemes. The authors suggest a robust facial feature detector based on a generalized symmetry interest operator. No special tuning is required if the face occupies 15-60% of the image. The operator was tested on a large face data base with a success rate of over 95%.>

An operator based on the intuitive motion of symmetry, which effectively locates interest points in real time and can be incorporated also in active visual systems, is introduced. The results of its operation agree with some psychophysical evidence concerning symmetry as well as evidence concerning fixation points. The operator can be applied successfully without prior knowledge of the world. Combining the operator with some preconceptions about the image is a powerful tool for feature detection in intricate natural scenes. The localization of faces and facial features in real time is demonstrated on detailed and noisy pictures.>

Neural networks consist of simple elements capable of summation and thresholding. The authors define a more general element, the task-oriented neuron or teuron, which can compute higher-order functions. They further define teuron networks and show two such networks that can be used as content addressable memories and as pattern classifiers. The first network is based on the Godel encoding scheme. It uses this scheme in order to memorize sequences of numbers. These sequences are then stored analogically. The second network uses binary encoding, i.e. a binary sequence is translated into its decimal equivalent and then stored analogically. The authors demonstrate the feasibility of implementing such networks. It is concluded that they can be implemented in such a way as to reduce cost, due to a reduction in the number of elements coupled with constancy of link values (synaptic weights).>

Many primate visual cortex architectures have a prominent feature responsible for the mixing of left and right eye visual data: ocular dominance columns represent thin (about 5-10 minutes of arc) strips of alternating left and right eye input to the brain. It is shown that such an architecture, when operated upon with a cepstral filter, provides a strong cue for binocular stereopsis. Specifically, the vector of binocular disparity may be easily identified in the output of the (columnar based) cepstral filter. This algorithm is illustrated with application to a random dot stereogram and to natural images. The authors suggest that this provides a fast algorithm for stereo segmentation, in a machine vision context. In a biological context, it may provide a computational rationale for the existence of columnar systems with regard to both ocular mixing and other visual modalities that have a columnar architecture.>

The authors discuss three algorithms related to the blending of a single scene from multiple frames acquired from a space-variant sensor. Given a series of space-variant contour-based scenes with different fixation points, they show how to fuse these into a single multiscan view, which incorporates the information present in the individual scans. They demonstrate an (attentional) algorithm which recursively examines the current knowledge of the scene in order best to choose the next fixation point in terms of focusing attention on regions of maximum boundary curvature. They discuss a simple metric for evaluating convergence over a scan path. This may be used to compare the performance of various attentional algorithms. They discuss their work in light of both machine and biological vision.>