One of the fundamental problems in machine learning is training high-quality neural network models using small quantities of data. In many situations, obtaining more data might not be possible due to an inherent rarity, physical lack of the samples, financial costs, or time constraints. Under these circumstances, standard training approaches lead to significant overfitting thus prompting the recent emergence of few-shot (or low-shot) learning, a problem setting where every category in a task has only a few examples available. In the visual domain, the initial effort has been directed at the task of image classification, however promising results in classification have triggered recent advances in few-shot object detection, image segmentation, and video classification.

Humans are very good at learning new concepts given only few instances and in this work we are aiming to mimic that ability. The task of few-shot image classification is not in itself yet well-studied as the state-of-the-art approaches depend highly on the number of available categories and per-category examples of a task. Existing few-shot multiclass methods assume that the test set is closed, meaning that only images truly belonging to the categories seen during training can be present in the test set, which is a heavy restriction limiting the real-world applications of these methods. To ensure that few-shot methods will be able to translate well to real-world applications we focus on few-shot tasks containing data coming from both known and unknown categories: one-class and multiclass open-set scenarios. To efficiently learn methods for these scenarios we concentrate on meta-learning (or learning-to-learn) approaches that aim to facilitate model training in few-shot learning tasks.

In this thesis, we introduce four separate meta-learning methods for few-shot one-class and few-shot multiclass open-set settings and set new state-of-the-art performance levels on these tasks. Firstly, we introduce a novel task of few-shot one-class image classification, where the training set consists of only a few examples from a single category with a goal of differentiating examples coming from this category from examples coming from other, unknown categories. We propose and analyze three novel methods to solve this problem through meta-learning: one trained in multiple stages to predict weights of an SVM classifier, the second one learning a separate feature space dedicated for one-class classification, and the third one trained end-to-end to dynamically generate one-class neural network classifiers. In the second part, we concentrate on the impact the number of per-category examples in the training set has on the existing methods in the few-shot multiclass closed-set classification task. We introduce a novel dynamic meta-learning solution to enhance the performance of existing few-shot learning methods while keeping the model fast and lightweight. Finally, we discuss approaches combining the analysis and methods introduced in previous parts of the thesis and apply them to address a recently introduced task of few-shot multiclass open-set classification. Proposed approaches significantly outperform the recently introduced state-of-the-art method on the few-shot multiclass open-set classification task.

Classifiers generally lack a mechanism to compute decision confidences. As humans, when we sense that the confidence for a decision is low, we either conduct additional actions to improve our confidence or dismiss the decision. While this reasoning is natural to us, it is currently missing in most common decision algorithms (i.e., classifiers) used in computer vision or machine learning. This limits the capability for a machine to take further actions to either improve a result or dismiss the decision. In this thesis, we design algorithms for estimating the confidence for decisions made by classifiers such as nearest-neighbor or support vector machines. We developed these algorithms leveraging the theory of extreme values. We use the statistical models that this theory provides for modeling the classifier's decision scores for correct and incorrect outcomes. Our proposed algorithms exploit these statistical models in order to compute a correctness belief: the probability that the classifier's decision is correct. In this work, we show how these beliefs can be used to filter bad classifications and to speed up robust estimations via sample and consensus algorithms, which are used in computer vision for estimating camera motions and for reconstructing the scene's 3D structure. Moreover, we show how these beliefs improve the classification accuracy of one-class support vector machines. In conclusion, we show that extreme value theory leads to powerful mechanisms that can predict the correctness of a classifier's decision.

With the widespread deployment of fast data connections and availability of a variety of sensors for different modalities, the potential of remote collaboration has greatly increased. While the now ubiquitous video conferencing applications take advantage of some of these capabilities, the use of video between remote users is limited to passively watching disjoint video feeds and provides no means for interaction with the remote environment. However, collaboration often involves sharing, exploring, referencing, or even manipulating the physical world, and thus tools should provide support for these interactions.

We suggest that augmented reality is an intuitive and user-friendly paradigm to communicate information about the physical environment, and that integration of computer vision and augmented reality facilitates more immersive and more direct interaction with the remote environment than what is possible with today's tools.

In this dissertation, we present contributions to realizing this vision on several levels. First, we describe a conceptual framework for unobtrusive mobile video-mediated communication in which the remote user can explore the live scene independent of the local user's current camera movement, and can communicate information by creating spatial annotations that are immediately visible to the local user in augmented reality. Second, we describe the design and implementation of several, increasingly more flexible and immersive user interfaces and system prototypes that implement this concept. Our systems do not require any preparation or instrumentation of the environment; instead, the physical scene is tracked and modeled incrementally using monocular computer vision. The emerging model then supports anchoring of annotations, virtual navigation, and synthesis of novel views of the scene. Third, we describe the design, execution and analysis of three user studies comparing our prototype implementations with more conventional interfaces and/or evaluating specific design elements. Study participants overwhelmingly preferred our technology, and their task performance was significantly better compared with a video-only interface, though no task performance difference was observed compared with a ``static marker'' interface. Last, we address a particular technical limitation of current monocular tracking and mapping systems which was found to be impeding and present a conceptual solution; namely, we describe a concept and proof-of-concept implementation for automatic model selection which allows tracking and modeling to cope with both parallax-inducing and rotation-only camera movements.

We suggest that our results demonstrate the maturity and usability of our systems, and, more importantly, the potential of our approach to improve video-mediated communication and broaden its applicability.

Structure-from-Motion (SfM) is a powerful tool for computing 3D reconstructions from images of a scene and has wide applications in computer vision, scene recognition, and augmented and virtual reality. Standard SfM pipelines make strict assumptions about the capturing devices in order to simplify the process for estimating camera geometry and 3D structure. Specifically, most methods require monocular cameras with known focal length calibration. When considering large-scale SfM from internet photo collections, EXIF calibrations cannot be used reliably. Further, the requirement of single camera systems limits the scalability of SfM.

This thesis proposes to remove these constraints by instead considering the collection of cameras as a distributed camera that encapsulates the image and geometric information of all cameras simultaneously. First, I provide full generalizations to the relative camera pose and absolute camera pose problems. These generalizations are more expressive and extend the traditional single-camera problems to distributed cameras, forming the basis for a novel hierarchical SfM pipeline that exhibits state-of-the-art performance on large-scale datasets. Second, I describe two efficient methods for estimating camera focal lengths for the distributed camera when calibration is not available. Finally, I show how removing these constraints enables a simpler, more scalable SfM pipeline that is capable of handling uncalibrated cameras at scale.