Learning Embeddings for Indexing, Retrieval, and Classification, with Applications to Object and Shape Recognition in Image Databases

Nearest neighbor retrieval is the task of identifying, given a database of objects and a query object, the objects in the database that are the most similar to the query. Retrieving nearest neighbors is a necessary component of many practical applications, in fields as diverse as computer vision, pattern recognition, multimedia databases, bioinformatics, and computer networks. At the same time, finding nearest neighbors accurately and efficiently can be challenging, especially when the database contains a large number of objects, and when the underlying distance measure is computationally expensive. This thesis proposes new methods for improving the efficiency and accuracy of nearest neighbor retrieval and classification in spaces with computationally expensive distance measures. The proposed methods are domain-independent, and can be applied in arbitrary spaces, including non-Euclidean and non-metric spaces. In this thesis particular emphasis is given to computer vision applications related to object and shape recognition, where expensive non-Euclidean distance measures are often needed to achieve high accuracy. The first contribution of this thesis is the BoostMap algorithm for embedding arbitrary spaces into a vector space with a computationally efficient distance measure. Using this approach, an approximate set of nearest neighbors can be retrieved efficiently - often orders of magnitude faster than retrieval using the exact distance measure in the original space. The BoostMap algorithm has two key distinguishing features with respect to existing embedding methods. First, embedding construction explicitly maximizes the amount of nearest neighbor information preserved by the embedding. Second, embedding construction is treated as a machine learning problem, in contrast to existing methods that are based on geometric considerations. The second contribution is a method for constructing query-sensitive distance measures for the purposes of nearest neighbor retrieval and classification. In high-dimensional spaces, query-sensitive distance measures allow for automatic selection of the dimensions that are the most informative for each specific query object. It is shown theoretically and experimentally that query-sensitivity increases the modeling power of embeddings, allowing embeddings to capture a larger amount of the nearest neighbor structure of the original space. The third contribution is a method for speeding up nearest neighbor classification by combining multiple embedding-based nearest neighbor classifiers in a cascade. In a cascade, computationally efficient classifiers are used to quickly classify easy cases, and classifiers that are more computationally expensive and also more accurate are only applied to objects that are harder to classify. An interesting property of the proposed cascade method is that, under certain conditions, classification time actually decreases as the size of the database increases, a behavior that is in stark contrast to the behavior of typical nearest neighbor classification systems. The proposed methods are evaluated experimentally in several different applications: hand shape recognition, off-line character recognition, online character recognition, and efficient retrieval of time series. In all datasets, the proposed methods lead to significant improvements in accuracy and efficiency compared to existing state-of-the-art methods. In some datasets, the general-purpose methods introduced in this thesis even outperform domain-specific methods that have been custom-designed for such datasets.

The What-And-Where Filter: A Spatial Mapping Neural Network for Object Recognition and Image Understanding

The What-and-Where filter forms part of a neural network architecture for spatial mapping, object recognition, and image understanding. The Where fllter responds to an image figure that has been separated from its background. It generates a spatial map whose cell activations simultaneously represent the position, orientation, ancl size of all tbe figures in a scene (where they are). This spatial map may he used to direct spatially localized attention to these image features. A multiscale array of oriented detectors, followed by competitve and interpolative interactions between position, orientation, and size scales, is used to define the Where filter. This analysis discloses several issues that need to be dealt with by a spatial mapping system that is based upon oriented filters, such as the role of cliff filters with and without normalization, the double peak problem of maximum orientation across size scale, and the different self-similar interpolation properties across orientation than across size scale. Several computationally efficient Where filters are proposed. The Where filter rnay be used for parallel transformation of multiple image figures into invariant representations that are insensitive to the figures' original position, orientation, and size. These invariant figural representations form part of a system devoted to attentive object learning and recognition (what it is). Unlike some alternative models where serial search for a target occurs, a What and Where representation can he used to rapidly search in parallel for a desired target in a scene. Such a representation can also be used to learn multidimensional representations of objects and their spatial relationships for purposes of image understanding. The What-and-Where filter is inspired by neurobiological data showing that a Where processing stream in the cerebral cortex is used for attentive spatial localization and orientation, whereas a What processing stream is used for attentive object learning and recognition.

Handsignals Recognition From Video Using 3D Motion Capture Data

Hand signals are commonly used in applications such as giving instructions to a pilot for airplane take off or direction of a crane operator by a foreman on the ground. A new algorithm for recognizing hand signals from a single camera is proposed. Typically, tracked 2D feature positions of hand signals are matched to 2D training images. In contrast, our approach matches the 2D feature positions to an archive of 3D motion capture sequences. The method avoids explicit reconstruction of the 3D articulated motion from 2D image features. Instead, the matching between the 2D and 3D sequence is done by backprojecting the 3D motion capture data onto 2D. Experiments demonstrate the effectiveness of the approach in an example application: recognizing six classes of basketball referee hand signals in video.

Modal Matching for Correspondence and Recognition

Modal matching is a new method for establishing correspondences and computing canonical descriptions. The method is based on the idea of describing objects in terms of generalized symmetries, as defined by each object's eigenmodes. The resulting modal description is used for object recognition and categorization, where shape similarities are expressed as the amounts of modal deformation energy needed to align the two objects. In general, modes provide a global-to-local ordering of shape deformation and thus allow for selecting which types of deformations are used in object alignment and comparison. In contrast to previous techniques, which required correspondence to be computed with an initial or prototype shape, modal matching utilizes a new type of finite element formulation that allows for an object's eigenmodes to be computed directly from available image information. This improved formulation provides greater generality and accuracy, and is applicable to data of any dimensionality. Correspondence results with 2-D contour and point feature data are shown, and recognition experiments with 2-D images of hand tools and airplanes are described.

Color Region Grouping and Shape Recognition with Deformable Models

A new deformable shape-based method for color region segmentation is described. The method includes two stages: over-segmentation using a traditional color region segmentation algorithm, followed by deformable model-based region merging via grouping and hypothesis selection. During the second stage, region merging and object identification are executed simultaneously. A statistical shape model is used to estimate the likelihood of region groupings and model hypotheses. The prior distribution on deformation parameters is precomputed using principal component analysis over a training set of region groupings. Once trained, the system autonomously segments deformed shapes from the background, while not merging them with similarly colored adjacent objects. Furthermore, the recovered parametric shape model can be used directly in object recognition and comparison. Experiments in segmentation and image retrieval are reported.

Indexing Methods for Efficient Multiclass Recognition

Many real world image analysis problems, such as face recognition and hand pose estimation, involve recognizing a large number of classes of objects or shapes. Large margin methods, such as AdaBoost and Support Vector Machines (SVMs), often provide competitive accuracy rates, but at the cost of evaluating a large number of binary classifiers, thus making it difficult to apply such methods when thousands or millions of classes need to be recognized. This thesis proposes a filter-and-refine framework, whereby, given a test pattern, a small number of candidate classes can be identified efficiently at the filter step, and computationally expensive large margin classifiers are used to evaluate these candidates at the refine step. Two different filtering methods are proposed, ClassMap and OVA-VS (One-vs.-All classification using Vector Search). ClassMap is an embedding-based method, works for both boosted classifiers and SVMs, and tends to map the patterns and their associated classes close to each other in a vector space. OVA-VS maps OVA classifiers and test patterns to vectors based on the weights and outputs of weak classifiers of the boosting scheme. At runtime, finding the strongest-responding OVA classifier becomes a classical vector search problem, where well-known methods can be used to gain efficiency. In our experiments, the proposed methods achieve significant speed-ups, in some cases up to two orders of magnitude, compared to exhaustive evaluation of all OVA classifiers. This was achieved in hand pose recognition and face recognition systems where the number of classes ranges from 535 to 48,600.

3D Trajectory Recovery for Tracking Multiple Objects and Trajectory Guided Recognition of Actions

A mechanism is proposed that integrates low-level (image processing), mid-level (recursive 3D trajectory estimation), and high-level (action recognition) processes. It is assumed that the system observes multiple moving objects via a single, uncalibrated video camera. A novel extended Kalman filter formulation is used in estimating the relative 3D motion trajectories up to a scale factor. The recursive estimation process provides a prediction and error measure that is exploited in higher-level stages of action recognition. Conversely, higher-level mechanisms provide feedback that allows the system to reliably segment and maintain the tracking of moving objects before, during, and after occlusion. The 3D trajectory, occlusion, and segmentation information are utilized in extracting stabilized views of the moving object. Trajectory-guided recognition (TGR) is proposed as a new and efficient method for adaptive classification of action. The TGR approach is demonstrated using "motion history images" that are then recognized via a mixture of Gaussian classifier. The system was tested in recognizing various dynamic human outdoor activities; e.g., running, walking, roller blading, and cycling. Experiments with synthetic data sets are used to evaluate stability of the trajectory estimator with respect to noise.

Recognition of Human Action Using Moment-Based Feature

The performance of different classification approaches is evaluated using a view-based approach for motion representation. The view-based approach uses computer vision and image processing techniques to register and process the video sequence. Two motion representations called Motion Energy Images and Motion History Image are then constructed. These representations collapse the temporal component in a way that no explicit temporal analysis or sequence matching is needed. Statistical descriptions are then computed using moment-based features and dimensionality reduction techniques. For these tests, we used 7 Hu moments, which are invariant to scale and translation. Principal Components Analysis is used to reduce the dimensionality of this representation. The system is trained using different subjects performing a set of examples of every action to be recognized. Given these samples, K-nearest neighbor, Gaussian, and Gaussian mixture classifiers are used to recognize new actions. Experiments are conducted using instances of eight human actions (i.e., eight classes) performed by seven different subjects. Comparisons in the performance among these classifiers under different conditions are analyzed and reported. Our main goals are to test this dimensionality-reduced representation of actions, and more importantly to use this representation to compare the advantages of different classification approaches in this recognition task.

Trajectory Guided Tracking and Recognition of Actions

A combined 2D, 3D approach is presented that allows for robust tracking of moving people and recognition of actions. It is assumed that the system observes multiple moving objects via a single, uncalibrated video camera. Low-level features are often insufficient for detection, segmentation, and tracking of non-rigid moving objects. Therefore, an improved mechanism is proposed that integrates low-level (image processing), mid-level (recursive 3D trajectory estimation), and high-level (action recognition) processes. A novel extended Kalman filter formulation is used in estimating the relative 3D motion trajectories up to a scale factor. The recursive estimation process provides a prediction and error measure that is exploited in higher-level stages of action recognition. Conversely, higher-level mechanisms provide feedback that allows the system to reliably segment and maintain the tracking of moving objects before, during, and after occlusion. The 3D trajectory, occlusion, and segmentation information are utilized in extracting stabilized views of the moving object that are then used as input to action recognition modules. Trajectory-guided recognition (TGR) is proposed as a new and efficient method for adaptive classification of action. The TGR approach is demonstrated using "motion history images" that are then recognized via a mixture-of-Gaussians classifier. The system was tested in recognizing various dynamic human outdoor activities: running, walking, roller blading, and cycling. Experiments with real and synthetic data sets are used to evaluate stability of the trajectory estimator with respect to noise.

Biologically Inspired Approaches to Automated Feature Extraction and Target Recognition

Ongoing research at Boston University has produced computational models of biological vision and learning that embody a growing corpus of scientific data and predictions. Vision models perform long-range grouping and figure/ground segmentation, and memory models create attentionally controlled recognition codes that intrinsically cornbine botton-up activation and top-down learned expectations. These two streams of research form the foundation of novel dynamically integrated systems for image understanding. Simulations using multispectral images illustrate road completion across occlusions in a cluttered scene and information fusion from incorrect labels that are simultaneously inconsistent and correct. The CNS Vision and Technology Labs (cns.bu.edulvisionlab and cns.bu.edu/techlab) are further integrating science and technology through analysis, testing, and development of cognitive and neural models for large-scale applications, complemented by software specification and code distribution.

CEDI: A Neural Model of Colour Vision, with Applications to Image Processing and Classification

Air Force Office of Scientific Research (F49620-01-1-0423); National Geospatial-Intelligence Agency (NMA 201-01-1-2016); National Science Foundation (SBE-035437, DEG-0221680); Office of Naval Research (N00014-01-1-0624)

Brain Learning and Recognition: The Large and the Small of It in Inferotemporal Cortex

Anterior inferotemporal cortex (ITa) plays a key role in visual object recognition. Recognition is tolerant to object position, size, and view changes, yet recent neurophysiological data show ITa cells with high object selectivity often have low position tolerance, and vice versa. A neural model learns to simulate both this tradeoff and ITa responses to image morphs using large-scale and small-scale IT cells whose population properties may support invariant recognition.

ART-EMAP: A Neural Network Architecture for Object Recognition by Evidence Accumulation

A new neural network architecture is introduced for the recognition of pattern classes after supervised and unsupervised learning. Applications include spatio-temporal image understanding and prediction and 3-D object recognition from a series of ambiguous 2-D views. The architecture, called ART-EMAP, achieves a synthesis of adaptive resonance theory (ART) and spatial and temporal evidence integration for dynamic predictive mapping (EMAP). ART-EMAP extends the capabilities of fuzzy ARTMAP in four incremental stages. Stage 1 introduces distributed pattern representation at a view category field. Stage 2 adds a decision criterion to the mapping between view and object categories, delaying identification of ambiguous objects when faced with a low confidence prediction. Stage 3 augments the system with a field where evidence accumulates in medium-term memory (MTM). Stage 4 adds an unsupervised learning process to fine-tune performance after the limited initial period of supervised network training. Each ART-EMAP stage is illustrated with a benchmark simulation example, using both noisy and noise-free data. A concluding set of simulations demonstrate ART-EMAP performance on a difficult 3-D object recognition problem.

A What-and-Where Neural Network for Invariant Image Preprocessing

A feedforward neural network for invariant image preprocessing is proposed that represents the position1 orientation and size of an image figure (where it is) in a multiplexed spatial map. This map is used to generate an invariant representation of the figure that is insensitive to position1 orientation, and size for purposes of pattern recognition (what it is). A multiscale array of oriented filters followed by competition between orientations and scales is used to define the Where filter.