Evidence is mounting that Convolutional Networks (ConvNets) are the most effective representation learning method for visual recognition tasks. In the common scenario, a ConvNet is trained on a large labeled dataset (source) and the feed-forward units activation of the trained network, at a certain layer of the network, is used as a generic representation of an input image for a task with relatively smaller training set (target). Recent studies have shown this form of representation transfer to be suitable for a wide range of target visual recognition tasks. This paper introduces and investigates several factors affecting the transferability of such representations. It includes parameters for training of the source ConvNet such as its architecture, distribution of the training data, etc. and also the parameters of feature extraction such as layer of the trained ConvNet, dimensionality reduction, etc. Then, by optimizing these factors, we show that significant improvements can be achieved on various (17) visual recognition tasks. We further show that these visual recognition tasks can be categorically ordered based on their similarity to the source task such that a correlation between the performance of tasks and their similarity to the source task w.r.t. the proposed factors is observed.

Discriminative latent variable models (LVM) are frequently applied to various visual recognition tasks. In these systems the latent (hidden) variables provide a formalism for modeling structured variation of visual features. Conventionally, latent variables are de- fined on the variation of the foreground (positive) class. In this work we augment LVMs to include negative latent variables corresponding to the background class. We formalize the scoring function of such a generalized LVM (GLVM). Then we discuss a framework for learning a model based on the GLVM scoring function. We theoretically showcase how some of the current visual recognition methods can benefit from this generalization. Finally, we experiment on a generalized form of Deformable Part Models with negative latent variables and show significant improvements on two different detection tasks.

We discuss the properties of a class of latent variable models that assumes each labeled sample is associated with a set of different features, with no prior knowledge of which feature is the most relevant feature to be used. Deformable-Part Models (DPM) can be seen as good examples of such models. These models are usually considered to be expensive to train and very sensitive to the initialization. In this paper, we focus on the learning of such models by introducing a topological framework and show how it is possible to both reduce the learning complexity and produce more robust decision boundaries. We will also argue how our framework can be used for producing robust decision boundaries without exploiting the dataset bias or relying on accurate annotations. To experimentally evaluate our method and compare with previously published frameworks, we focus on the problem of image classification with object localization. In this problem, the correct location of the objects is unknown, during both training and testing stages, and is considered as a latent variable.

This paper demonstrates a system for automatic detection of visual attention and identification of salient items at exhibitions (e.g. museum or an auction). The method is offline and is done on a video captured by a head mounted camera. Towards the estimation of attention, we define the notions of "saliency" and "interestingness" for an exhibition items. Our method is a combination of multiple state of the art techniques from different vision tasks such as tracking, image matching and retrieval. Many experiments are conducted to evaluate multiple aspects of our method. The method has proven to be robust to image blur, occlusion, truncation, and dimness. The experiments shows strong performance for the tasks of matching items, estimating focus frames and detecting salient and interesting items. This can be useful to the commercial vendors and museum curators and help them to understand which items are appealing more to the visitors.

In this paper, we discuss the properties of a class of latent variable models that assumes each labeled sample is associated with set of different features, with no prior knowledge of which feature is the most relevant feature to be used. Deformable-Part Models (DPM) can be seen as good example of such models. While Latent SVM framework (LSVM) has proven to be an efficient tool for solving these models, we will argue that the solution found by this tool is very sensitive to the initialization. To decrease this dependency, we propose a novel clustering procedure, for these problems, to find cluster centers that are shared by several sample sets while ignoring the rest of the cluster centers. As we will show, these cluster centers will provide a robust initialization for the LSVM framework.

In this paper, we propose a framework to gradually improve the quality of an already existing image descriptor. The descriptor used in this paper \cite{Afkham:wf} uses the response of a series of discriminative components for summarizing each image. As we will show, this descriptor has an ideal form in which all categories become linearly separable. While, reaching this form is not feasible, we will argue by replacing a small fraction of the components, it is possible to obtain a descriptor which is, on average, closer to this ideal form. To do so, we identify which components do not contribute to the quality of the descriptor and replace them with more robust components. Here, a joint feature selection method is used to find more robust components. As our experiments show, this change directly reflects in the capability of the resulting descriptor in discriminating between different categories.

This thesis is mostly about supervised visual recognition problems. Based on a general definition of categories, the contents are divided into two parts: one which models categories and one which is not category based. We are interested in data driven solutions for both kinds of problems.In the category-free part, we study novelty detection in temporal and spatial domains as a category-free recognition problem. Using data driven models, we demonstrate that based on a few reference exemplars, our methods are able to detect novelties in ego-motions of people, and changes in the static environments surrounding them.In the category level part, we study object recognition. We consider both object category classification and localization, and propose scalable data driven approaches for both problems. A mixture of parametric classifiers, initialized with a sophisticated clustering of the training data, is demonstrated to adapt to the data better than various baselines such as the same model initialized with less subtly designed procedures. A nonparametric large margin classifier is introduced and demonstrated to have a multitude of advantages in comparison to its competitors: better training and testing time costs, the ability to make use of indefinite/invariant and deformable similarity measures, and adaptive complexity are the main features of the proposed model.We also propose a rather realistic model of recognition problems, which quantifies the interplay between representations, classifiers, and recognition performances. Based on data-describing measures which are aggregates of pairwise similarities of the training data, our model characterizes and describes the distributions of training exemplars. The measures are shown to capture many aspects of the difficulty of categorization problems and correlate significantly to the observed recognition performances. Utilizing these measures, the model predicts the performance of particular classifiers on distributions similar to the training data. These predictions, when compared to the test performance of the classifiers on the test sets, are reasonably accurate.We discuss various aspects of visual recognition problems: what is the interplay between representations and classification tasks, how can different models better adapt to the training data, etc. We describe and analyze the aforementioned methods that are designed to tackle different visual recognition problems, but share one common characteristic: being data driven.

We consider the problem of automatically estimating the 3D pose of humans from images, taken from multiple calibrated views. We show that it is possible and tractable to extend the pictorial structures framework, popular for 2D pose estimation, to 3D. We discuss how to use this framework to impose view, skeleton, joint angle and intersection constraints in 3D. The 3D pictorial structures are evaluated on multiple view data from a professional football game. The evaluation is focused on computational tractability, but we also demonstrate how a simple 2D part detector can be plugged into the framework.

Most object classes share a considerable amount of local appearance and often only a small number of features are discriminative. The traditional approach to represent an object is based on a summarization of the local characteristics by counting the number of feature occurrences. In this paper we propose the use of a recently developed technique for summarizations that, rather than looking into the quantity of features, encodes their quality to learn a description of an object. Our approach is based on extracting and aggregating only the essential characteristics of an object class for a task. We show how the proposed method significantly improves on previous work in 3D object categorization. We discuss the benefits of the method in other scenarios such as robot grasping. We provide extensive quantitative and qualitative experiments comparing our approach to the state of the art to justify the described approach.

We propose a system for the automatic segmentation of novelties from the background in scenarios where multiple images of the same environment are available e.g. obtained by wearable visual cameras. Our method finds the pixels in a query image corresponding to the underlying background environment by comparing it to reference images of the same scene. This is achieved despite the fact that all the images may have different viewpoints, significantly different illumination conditions and contain different objects - cars, people, bicycles, etc. - occluding the background. We estimate the probability of each pixel, in the query image, belonging to the background by computing its appearance inconsistency to the multiple reference images. We then, produce multiple segmentations of the query image using an iterated graph cuts algorithm, initializing from these estimated probabilities and consecutively combine these segmentations to come up with a final segmentation of the background. Detection of the background in turn highlights the novel pixels. We demonstrate the effectiveness of our approach on a challenging outdoors data set.

This paper demonstrates a system for the automatic extraction of novelty in images captured from a small video camera attached to a subject's chest, replicating his visual perspective, while performing activities which are repeated daily. Novelty is detected when a (sub)sequence cannot be registered to previously stored sequences captured while performing the same daily activity. Sequence registration is performed by measuring appearance and geometric similarity of individual frames and exploiting the invariant temporal order of the activity. Experimental results demonstrate that this is a robust way to detect novelties induced by variations in the wearer's ego-motion such as stopping and talking to a person. This is an essentially new and generic way of automatically extracting information of interest to the camera wearer and can be used as input to a system for life logging or memory support.

In this paper we address the problem of using boosting (e.g. AdaBoost [7]) to classify a target class with significant intra-class variation against a large background class. This situation occurs for example when we want to recognize a visual object class against all other image patches. The boosting algorithm produces a strong classifier, which is a linear combination of weak classifiers. We observe that we often have sets of weak classifiers that individually fire on many examples of the target class but never fire together on those examples (i.e. their outputs are anti-correlated on the target class). Motivated by this observation we suggest a family of derived weak classifiers, termed gated classifiers, that suppress such combinations of weak classifiers. Gated classifiers can be used on top of any original weak learner. We run experiments on two popular datasets, showing that our method reduces the required number of weak classifiers by almost an order of magnitude, which in turn yields faster detectors. We experiment on synthetic data showing that gated classifiers enables more complex distributions to be represented. We hope that gated classifiers will extend the usefulness of boosted classifier cascades [29].

In this paper we introduce a new kind of feature - the multi-local feature, so named as each one is a collection of local features, such as oriented edgels, in a very specific spatial arrangement. A multi-local feature has the ability to capture underlying constant shape properties of exemplars from an object class. Thus it is particularly suited to representing and detecting visual classes that lack distinctive local structures and are mainly defined by their global shape. We present algorithms to automatically learn an ensemble of these features to represent an object class from weakly labelled training images of that class, as well as procedures to detect these features efficiently in novel images. The power of multi-local features is demonstrated by using the ensemble in a simple voting scheme to perform object category detection on a standard database. Despite its simplicity, this scheme yields detection rates matching state-of-the-art object detection systems.

Multi-target tracking requires locating the targets and labeling their identities. The latter is a challenge when many targets, with indistinct appearances, frequently occlude one another, as in football and surveillance tracking. We present an approach to solving this labeling problem. When isolated, a target can be tracked and its identity maintained. While, if targets interact this is not always the case. This paper assumes a track graph exists, denoting when targets are isolated and describing how they interact. Measures of similarity between isolated tracks are defined. The goal is to associate the identities of the isolated tracks, by exploiting the graph constraints and similarity measures. We formulate this as a Bayesian network inference problem, allowing us to use standard message propagation to find the most probable set of paths in an efficient way. The high complexity inevitable in large problems is gracefully reduced by removing dependency links between tracks. We apply the method to a 10 min sequence of an international football game and compare results to ground truth.

We present an algorithm for 3D reconstruction of dynamic articulated structures, such as humans, from uncalibrated multiple views. The reconstruction exploits constraints associated with a dynamic articulated structure, specifically the conservation over time of length between rotational joints. These constraints admit metric reconstruction from at least two different images in each of two uncalibrated parallel projection cameras. The algorithm is based on a stratified approach, starting with affine reconstruction from factorization, followed by rectification to metric structure using the articulated structure constraints. The exploitation of these specific constraints allows reconstruction and self-calibration with fewer feature paints and views compared to standard self-calibration. The method is extended to pairs of cameras that are zooming, Where calibration of the cameras allows compensation for the changing scale factor in a scaled orthographic camera. Results are presented in the form of stick figures and animated 3D reconstructions using pairs of sequences from broadcast television. The technique shows promise as a means of creating 3D animations of dynamic activities such as sports events.

Abstract Cross ratios of image points are viewpoint invariants if the points are coplanar in 3D. If points are in general position in 3D, cross ratios will still be unaffected by intrinsic camera parameters and camera rotation, assuming linear imaging. The variation of cross ratios with viewpoint is therefore due entirely to the shift in relative position of the camera to the points in 3D. By inverting this relationship, we can compute camera position relative to a model of the environment using indexing methods. This computation is then independent of camera calibration.

An algorithm is presented for computing a decomposition of planar shapes into convex subparts represented by ellipses. The method is invariant to projective transformations of the shape, and thus the ellipses can be used for matching and definition of invariants in the same way as points and lines. The method works for arbitrary planar shapes admitting at least four distinct tangents and is based on finding ellipses with four points of contact to the given shape. The cross ratio computed from the four points on the ellipse can then be used as a projectively invariant index. For a given shape, each pair of ellipses can be used to compute two independent projective invariants. The set of invariants computed for each ellipse pair can be used as indexes to a hash table from which a model hypothesis can be generated.>

We present an algorithm for the extraction of a compact and accurate representation of the displacement field in moving image sequences. The algorithm works in two steps. First motion is estimated for contour segments assuming they move rigidly in the image plane. Thereafter the displacement field between contours is obtained by interpolation.