Preview

PDF, English Download (81MB) | Terms of use

Abstract

The quality of a Computer Vision system is proportional to the rigor of data representation it is built upon. Learning expressive representations of images is therefore the centerpiece to almost every computer vision application, including image search, object detection and classification, human re-identification, object tracking, pose understanding, image-to-image translation, and embodied agent navigation to name a few. Deep Neural Networks are most often seen among the modern methods of representation learning. The limitation is, however, that deep representation learning methods require extremely large amounts of manually labeled data for training. Clearly, annotating vast amounts of images for various environments is infeasible due to cost and time constraints. This requirement of obtaining labeled data is a prime restriction regarding pace of the development of visual recognition systems.

In order to cope with the exponentially growing amounts of visual data generated daily, machine learning algorithms have to at least strive to scale at a similar rate. The second challenge consists in the learned representations having to generalize to novel objects, classes, environments and tasks in order to accommodate to the diversity of the visual world. Despite the evergrowing number of recent publications tangentially addressing the topic of learning generalizable representations, efficient generalization is yet to be achieved. This dissertation attempts to tackle the problem of learning visual representations that can generalize to novel settings while requiring few labeled examples.

In this research, we study the limitations of the existing supervised representation learning approaches and propose a framework that improves the generalization of learned features by exploiting visual similarities between images which are not captured by provided manual annotations. Furthermore, to mitigate the common requirement of large scale manually annotated datasets, we propose several approaches that can learn expressive representations without human-attributed labels, in a self-supervised fashion, by grouping highly-similar samples into surrogate classes based on progressively learned representations.

The development of computer vision as science is preconditioned upon the seamless ability of a machine to record and disentangle pictures' attributes that were expected to only be conceived by humans. As such, particular interest was dedicated to the ability to analyze the means of artistic expression and style which depicts a more complex task than merely breaking an image down to colors and pixels. The ultimate test for this ability is the task of style transfer which involves altering the style of an image while keeping its content. An effective solution of style transfer requires learning such image representation which would allow disentangling image style and its content. Moreover, particular artistic styles come with idiosyncrasies that affect which content details should be preserved and which discarded. Another pitfall here is that it is impossible to get pixel-wise annotations of style and how the style should be altered. We address this problem by proposing an unsupervised approach that enables encoding the image content in such a way that is required by a particular style. The proposed approach exchanges the style of an input image by first extracting the content representation in a style-aware way and then rendering it in a new style using a style-specific decoder network, achieving compelling results in image and video stylization.

Finally, we combine supervised and self-supervised representation learning techniques for the task of human and animals pose understanding. The proposed method enables transfer of the representation learned for recognition of human poses to proximal mammal species without using labeled animal images. This approach is not limited to dense pose estimation and could potentially enable autonomous agents from robots to self-driving cars to retrain themselves and adapt to novel environments based on learning from previous experiences.