Preview

PDF, English Download (18MB) | Terms of use

Abstract

The visual world is a conglomeration of objects, scenes, motion, and much more. As humans, we look at the world through our eyes, but we understand it by using our brains. From a young age, humans learn to recognize objects by association, meaning that we link an object or action to the most similar one in our memory to make sense of it. Within the field of Artificial Intelligence, Computer Vision gives machines the ability to see. While digital cameras provide eyes to the machine, Computer Vision develops its brain. To that purpose, Deep Learning has emerged as a very successful tool. This method allows machines to learn solutions to problems directly from the data. On the basis of Deep Learning, computers nowadays can also learn to interpret the visual world. However, the process of learning in machines is very different from ours. In Deep Learning, images and videos are grouped into predefined, artificial categories. However, describing a group of objects, or actions, with a single integer (category) disregards most of its characteristics and pair-wise relationships. To circumvent this, we propose to expand the categorical model by using visual similarity which better mirrors the human approach. Deep Learning requires a large set of manually annotated samples, that form the training set. Retrieving training samples is easy given the endless amount of images and videos available on the internet. However, this also requires manual annotations, which are very costly and laborious to obtain and thus a major bottleneck in modern computer vision. In this thesis, we investigate visual similarity methods to solve image and video classification. In particular, we search for a solution where human super- vision is marginal. We focus on Zero-Shot Learning (ZSL), where only a subset of categories are manually annotated. After studying existing methods in the field, we identify common limitations and propose methods to tackle them. In particular, ZSL image classification is trained using only discriminative supervi- sion, i.e. predefined categories, while ignoring other descriptive characteristics. To tackle this, we propose a new approach to learn shared features, i.e. non- discriminative, thus descriptive characteristics, which improves existing methods by a large margin. However, while ZSL has shown great potential for the task of image classification, for example in case of face recognition, it has performed poorly for video classification. We identify the reasons for the lack of growth in the field and provide a new, powerful baseline. Unfortunately, even if ZSL requires only partial labeled data, it still needs supervision during training. For that reason, we also investigate purely unsuper- vised methods. A successful paradigm is self-supervision: the model is trained using a surrogate task where supervision is automatically provided. The key to self-supervision is the ability of deep learning to transfer the knowledge learned from one task to a new task. The more similar the two tasks are, the more effective the transfer is. Similar to our work on ZSL, we also studied the com- mon limitations of existing self-supervision approaches and proposed a method to overcome them. To improve self-supervised learning, we propose a policy network which controls the parameters of the surrogate task and is trained through reinforcement learning. Finally, we present a real-life application where utilizing visual similarity with limited supervision provides a better solution compared to existing parametric approaches. We analyze the behavior of motor-impaired rodents during a single repeating action for which our method provides an objective similarity of behav- ior, facilitating comparisons across animal subjects and time during recovery.