Computational Models of Object Recognition in Cortex: A Review

Understanding how biological visual systems perform object recognition is one of the ultimate goals in computational neuroscience. Among the biological models of recognition the main distinctions are between feedforward and feedback and between object-centered and view-centered. From a computational viewpoint the different recognition tasks - for instance categorization and identification - are very similar, representing different trade-offs between specificity and invariance. Thus the different tasks do not strictly require different classes of models. The focus of the review is on feedforward, view-based models that are supported by psychophysical and physiological data.

Stimulus Simplification and Object Representation: A Modeling Study

Tsunoda et al. (2001) recently studied the nature of object representation in monkey inferotemporal cortex using a combination of optical imaging and extracellular recordings. In particular, they examined IT neuron responses to complex natural objects and "simplified" versions thereof. In that study, in 42% of the cases, optical imaging revealed a decrease in the number of activation patches in IT as stimuli were "simplified". However, in 58% of the cases, "simplification" of the stimuli actually led to the appearance of additional activation patches in IT. Based on these results, the authors propose a scheme in which an object is represented by combinations of active and inactive columns coding for individual features. We examine the patterns of activation caused by the same stimuli as used by Tsunoda et al. in our model of object recognition in cortex (Riesenhuber 99). We find that object-tuned units can show a pattern of appearance and disappearance of features identical to the experiment. Thus, the data of Tsunoda et al. appear to be in quantitative agreement with a simple object-based representation in which an object's identity is coded by its similarities to reference objects. Moreover, the agreement of simulations and experiment suggests that the simplification procedure used by Tsunoda (2001) is not necessarily an accurate method to determine neuronal tuning.

Relative Affine Structure: Canonical Model for 3D from 2D Geometry and Applications

We propose an affine framework for perspective views, captured by a single extremely simple equation based on a viewer-centered invariant we call "relative affine structure". Via a number of corollaries of our main results we show that our framework unifies previous work --- including Euclidean, projective and affine --- in a natural and simple way, and introduces new, extremely simple, algorithms for the tasks of reconstruction from multiple views, recognition by alignment, and certain image coding applications.

Context-Based Vision System for Place and Object Recognition

While navigating in an environment, a vision system has to be able to recognize where it is and what the main objects in the scene are. In this paper we present a context-based vision system for place and object recognition. The goal is to identify familiar locations (e.g., office 610, conference room 941, Main Street), to categorize new environments (office, corridor, street) and to use that information to provide contextual priors for object recognition (e.g., table, chair, car, computer). We present a low-dimensional global image representation that provides relevant information for place recognition and categorization, and how such contextual information introduces strong priors that simplify object recognition. We have trained the system to recognize over 60 locations (indoors and outdoors) and to suggest the presence and locations of more than 20 different object types. The algorithm has been integrated into a mobile system that provides real-time feedback to the user.

Automatically Recovering Geometry and Texture from Large Sets of Calibrated Images

Three-dimensional models which contain both geometry and texture have numerous applications such as urban planning, physical simulation, and virtual environments. A major focus of computer vision (and recently graphics) research is the automatic recovery of three-dimensional models from two-dimensional images. After many years of research this goal is yet to be achieved. Most practical modeling systems require substantial human input and unlike automatic systems are not scalable. This thesis presents a novel method for automatically recovering dense surface patches using large sets (1000's) of calibrated images taken from arbitrary positions within the scene. Physical instruments, such as Global Positioning System (GPS), inertial sensors, and inclinometers, are used to estimate the position and orientation of each image. Essentially, the problem is to find corresponding points in each of the images. Once a correspondence has been established, calculating its three-dimensional position is simply a matter of geometry. Long baseline images improve the accuracy. Short baseline images and the large number of images greatly simplifies the correspondence problem. The initial stage of the algorithm is completely local and scales linearly with the number of images. Subsequent stages are global in nature, exploit geometric constraints, and scale quadratically with the complexity of the underlying scene. We describe techniques for: 1) detecting and localizing surface patches; 2) refining camera calibration estimates and rejecting false positive surfels; and 3) grouping surface patches into surfaces and growing the surface along a two-dimensional manifold. We also discuss a method for producing high quality, textured three-dimensional models from these surfaces. Some of the most important characteristics of this approach are that it: 1) uses and refines noisy calibration estimates; 2) compensates for large variations in illumination; 3) tolerates significant soft occlusion (e.g. tree branches); and 4) associates, at a fundamental level, an estimated normal (i.e. no frontal-planar assumption) and texture with each surface patch.

Geometry and Photometry in 3D Visual Recognition

The report addresses the problem of visual recognition under two sources of variability: geometric and photometric. The geometric deals with the relation between 3D objects and their views under orthographic and perspective projection. The photometric deals with the relation between 3D matte objects and their images under changing illumination conditions. Taken together, an alignment-based method is presented for recognizing objects viewed from arbitrary viewing positions and illuminated by arbitrary settings of light sources.

Sharing visual features for multiclass and multiview object detection

We consider the problem of detecting a large number of different classes of objects in cluttered scenes. Traditional approaches require applying a battery of different classifiers to the image, at multiple locations and scales. This can be slow and can require a lot of training data, since each classifier requires the computation of many different image features. In particular, for independently trained detectors, the (run-time) computational complexity, and the (training-time) sample complexity, scales linearly with the number of classes to be detected. It seems unlikely that such an approach will scale up to allow recognition of hundreds or thousands of objects. We present a multi-class boosting procedure (joint boosting) that reduces the computational and sample complexity, by finding common features that can be shared across the classes (and/or views). The detectors for each class are trained jointly, rather than independently. For a given performance level, the total number of features required, and therefore the computational cost, is observed to scale approximately logarithmically with the number of classes. The features selected jointly are closer to edges and generic features typical of many natural structures instead of finding specific object parts. Those generic features generalize better and reduce considerably the computational cost of an algorithm for multi-class object detection.

Viewpoint-Specific Representations in Three-Dimensional Object Recognition

We report a series of psychophysical experiments that explore different aspects of the problem of object representation and recognition in human vision. Contrary to the paradigmatic view which holds that the representations are three-dimensional and object-centered, the results consistently support the notion of view-specific representations that include at most partial depth information. In simulated experiments that involved the same stimuli shown to the human subjects, computational models built around two-dimensional multiple-view representations replicated our main psychophysical results, including patterns of generalization errors and the time course of perceptual learning.

Linear Object Classes and Image Synthesis from a Single Example Image

The need to generate new views of a 3D object from a single real image arises in several fields, including graphics and object recognition. While the traditional approach relies on the use of 3D models, we have recently introduced techniques that are applicable under restricted conditions but simpler. The approach exploits image transformations that are specific to the relevant object class and learnable from example views of other "prototypical" objects of the same class. In this paper, we introduce such a new technique by extending the notion of linear class first proposed by Poggio and Vetter. For linear object classes it is shown that linear transformations can be learned exactly from a basis set of 2D prototypical views. We demonstrate the approach on artificial objects and then show preliminary evidence that the technique can effectively "rotate" high- resolution face images from a single 2D view.

Error Detection and Recovery for Robot Motion Planning with Uncertainty

Robots must plan and execute tasks in the presence of uncertainty. Uncertainty arises from sensing errors, control errors, and uncertainty in the geometry of the environment. The last, which is called model error, has received little previous attention. We present a framework for computing motion strategies that are guaranteed to succeed in the presence of all three kinds of uncertainty. The motion strategies comprise sensor-based gross motions, compliant motions, and simple pushing motions.

Attentional Selection in Object Recognition

A key problem in object recognition is selection, namely, the problem of identifying regions in an image within which to start the recognition process, ideally by isolating regions that are likely to come from a single object. Such a selection mechanism has been found to be crucial in reducing the combinatorial search involved in the matching stage of object recognition. Even though selection is of help in recognition, it has largely remained unsolved because of the difficulty in isolating regions belonging to objects under complex imaging conditions involving occlusions, changing illumination, and object appearances. This thesis presents a novel approach to the selection problem by proposing a computational model of visual attentional selection as a paradigm for selection in recognition. In particular, it proposes two modes of attentional selection, namely, attracted and pay attention modes as being appropriate for data and model-driven selection in recognition. An implementation of this model has led to new ways of extracting color, texture and line group information in images, and their subsequent use in isolating areas of the scene likely to contain the model object. Among the specific results in this thesis are: a method of specifying color by perceptual color categories for fast color region segmentation and color-based localization of objects, and a result showing that the recognition of texture patterns on model objects is possible under changes in orientation and occlusions without detailed segmentation. The thesis also presents an evaluation of the proposed model by integrating with a 3D from 2D object recognition system and recording the improvement in performance. These results indicate that attentional selection can significantly overcome the computational bottleneck in object recognition, both due to a reduction in the number of features, and due to a reduction in the number of matches during recognition using the information derived during selection. Finally, these studies have revealed a surprising use of selection, namely, in the partial solution of the pose of a 3D object.

Reflectance Map Techniques for Analyzing Surface Defects in Metal Castings

This report explores the relation between image intensity and object shape. It is shown that image intensity is related to surface orientation and that a variation in image intensity is related to surface curvature. Computational methods are developed which use the measured intensity variation across surfaces of smooth objects to determine surface orientation. In general, surface orientation is not determined locally by the intensity value recorded at each image point. Tools are needed to explore the problem of determining surface orientation from image intensity. The notion of gradient space , popularized by Huffman and Mackworth, is used to represent surface orientation. The notion of a reflectance map, originated by Horn, is used to represent the relation between surface orientation image intensity. The image Hessian is defined and used to represent surface curvature. Properties of surface curvature are expressed as constraints on possible surface orientations corresponding to a given image point. Methods are presented which embed assumptions about surface curvature in algorithms for determining surface orientation from the intensities recorded in a single view. If additional images of the same object are obtained by varying the direction of incident illumination, then surface orientation is determined locally by the intensity values recorded at each image point. This fact is exploited in a new technique called photometric stereo. The visual inspection of surface defects in metal castings is considered. Two casting applications are discussed. The first is the precision investment casting of turbine blades and vanes for aircraft jet engines. In this application, grain size is an important process variable. The existing industry standard for estimating the average grain size of metals is implemented and demonstrated on a sample turbine vane. Grain size can be computed form the measurements obtained in an image, once the foreshortening effects of surface curvature are accounted for. The second is the green sand mold casting of shuttle eyes for textile looms. Here, physical constraints inherent to the casting process translate into these constraints, it is necessary to interpret features of intensity as features of object shape. Both applications demonstrate that successful visual inspection requires the ability to interpret observed changes in intensity in the context of surface topography. The theoretical tools developed in this report provide a framework for this interpretation.

On Motion Planning with Uncertainty

Robots must successfully plan and execute tasks in the presence of uncertainty. Uncertainty arises from errors in modeling, sensing, and control. Planning in the presence of uncertainty constitutes one facet of the general motion planning problem in robotics. This problem is concerned with the automatic synthesis of motion strategies from high level task specification and geometric models of environments. In order to develop successful motion strategies, it is necessary to understand the effect of uncertainty on the geometry of object interactions. Object interactions, both static and dynamic, may be represented in geometrical terms. This thesis investigates geometrical tools for modeling and overcoming uncertainty. The thesis describes an algorithm for computing backprojections o desired task configurations. Task goals and motion states are specified in terms of a moving object's configuration space. Backprojections specify regions in configuration space from which particular motions are guaranteed to accomplish a desired task. The backprojection algorithm considers surfaces in configuration space that facilitate sliding towards the goal, while avoiding surfaces on which motions may prematurely halt. In executing a motion for a backprojection region, a plan executor must be able to recognize that a desired task has been accomplished. Since sensors are subject to uncertainty, recognition of task success is not always possible. The thesis considers the structure of backprojection regions and of task goals that ensures goal recognizability. The thesis also develops a representation of friction in configuration space, in terms of a friction cone analogous to the real space friction cone. The friction cone provides the backprojection algorithm with a geometrical tool for determining points at which motions may halt.