Not all categories are created equal in object recognition. Fine-grained visual categorization (FGVC) is a branch of visual object recognition that aims to distinguish subordinate categories within a basic-level category. Examples include classifying an image of a bird into specific species like "Western Gull" or "California Gull". Such subordinate categories exhibit characteristics like small inter-category variation and large intra-class variation, making distinguishing them extremely difficult. To address such challenges, an algorithm should be able to focus on object parts and be invariant to object pose. Like many other computer vision tasks, FGVC has witnessed phenomenal advancement following the resurgence of deep neural networks. However, the proposed deep models are usually treated as black boxes. Network interpretation and understanding aims to unveil the features learned by neural networks and explain the reason behind network decisions. It is not only a necessary component for building trust between humans and algorithms, but also an essential step towards continuous improvement in this field. This dissertation is a collection of papers that contribute to FGVC and neural network interpretation and understanding. Our first contribution is an algorithm named Pose and Appearance Integration for Recognizing Subcategories (PAIRS) which performs pose estimation and generates a unified object representation as the concatenation of pose-aligned region features. As the second contribution, we propose the task of semantic network interpretation. For filter interpretation, we represent the concepts a filter detects using an attribute probability density function. We propose the task of semantic attribution using textual summarization that generates an explanatory sentence consisting of the most important visual attributes for decision-making, as found by a general Bayesian inference algorithm. Pooling has been a key component in convolutional neural networks and is of special interest in FGVC. Our third contribution is an empirical and experimental study towards a thorough yet intuitive understanding and extensive benchmark of popular pooling approaches. Our fourth contribution is a novel LMPNet for weakly-supervised keypoint discovery. A novel leaky max pooling layer is proposed to explicitly encourages sparse feature maps to be learned. A learnable clustering layer is proposed to group the keypoint proposals into final keypoint predictions. 2020 marks the 10th year since the beginning of fine-grained visual categorization. It is of great importance to summarize the representative works in this domain. Our last contribution is a comprehensive survey of FGVC containing nearly 200 relevant papers that cover 7 common themes.

The ability for unmanned autonomous aerial vehicles (UAV) to localize themselves in an environment is fundamental for them to be able to function, even if they do not have access to a global positioning system. Recently, with the success of deep learning in vision based tasks, there have been some proposed methods for absolute geolocalization using vison based deep learning with satellite and UAV images. Most of these are only tested in urban environments, which begs the question: How well do they work in non-urban areas like forests and fields? One drawback of deep learning is that models are often regarded as black boxes, as it is hard to know why the models make the predictions they do, i.e. what information is important and is used for the prediction. To solve this, several neural network interpretation methods have been developed. These methods provide explanations so that we may understand these models better. This thesis investigates the localization accuracy of one geolocalization method in both urban and non-urban environments as well as applies neural network interpretation in order to see if it can explain the potential difference in localization accuracy of the method in these different environments. The results show that the method performs best in urban environments, getting a mean absolute horizontal error of 38.30m and a mean absolute vertical error of 16.77m, while it performed significantly worse in non-urban environments, getting a mean absolute horizontal error of 68.11m and a mean absolute vertical error 22.83m. Further, the results show that if the satellite images and images from the unmanned aerial vehicle are collected during different seasons of the year, the localization accuracy is even worse, resulting in a mean absolute horizontal error of 86.91m and a mean absolute vertical error of 23.05m. The neural network interpretation did not aid in providing an explanation for why the method performs worse in non-urban environments and is not suitable for this kind of problem. / Obemannade autonoma luftburna fordons (UAV) förmåga att lokaliera sig själva är fundamental för att de ska fungera, även om de inte har tillgång till globala positioneringssystem. Med den nyliga framgången hos djupinlärning applicerat på visuella problem har det kommit metoder för absolut geolokalisering med visuell djupinlärning med satellit- och UAV-bilder. De flesta av dessa metoder har bara blivit testade i stadsmiljöer, vilket leder till frågan: Hur väl fungerar dessa metoder i icke-urbana områden som fält och skogar? En av nackdelarna med djupinlärning är att dessa modeller ofta ses som svarta lådor eftersom det är svårt att veta varför modellerna gör de gissningar de gör, alltså vilken information som är viktig och används för gissningen. För att lösa detta har flera metoder för att tolka neurala nätverk utvecklats. Dessa metoder ger förklaringar så att vi kan förstå dessa modeller bättre. Denna uppsats undersöker lokaliseringsprecisionen hos en geolokaliseringsmetod i både urbana och icke-urbana miljöer och applicerar även en tolkningsmetod för neurala nätverk för att se ifall den kan förklara den potentialla skillnaden i precision hos metoden i dessa olika miljöer. Resultaten visar att metoden fungerar bäst i urbana miljöer där den får ett genomsnittligt absolut horisontellt lokaliseringsfel på 38.30m och ett genomsnittligt absolut vertikalt fel på 16.77m medan den presterade signifikant sämre i icke-urbana miljöer där den fick ett genomsnittligt absolut horisontellt lokaliseringsfel på 68.11m och ett genomsnittligt absolut vertikalt fel på 22.83m. Vidare visar resultaten att om satellitbilderna och UAV-bilderna är tagna från olika årstider blir lokaliseringsprecisionen ännu sämre, där metoden får genomsnittligt absolut horisontellt lokaliseringsfel på 86.91m och ett genomsnittligt absolut vertikalt fel på 23.05m. Tolkningsmetoden hjälpte inte i att förklara varför metoden fungerar sämre i icke-urbana miljöer och är inte passande att använda för denna sortens problem.