Abstract: Put succinctly, the bag-of-words based image representation is the most successful approach for object and scene recognition. Within the bag-of-words framework the optimal fusion of multiple cues, such as shape, texture and color, still remains an active research domain. There exist two main approaches to combine color and shape information within the bag-of-words framework. The first approach called, early fusion, fuses color and shape at the feature level as a result of which a joint colorshape vocabulary is produced. The second approach, called late fusion, concatenates histogram representation of both color and shape, obtained independently. In the first part of this thesis, we analyze the theoretical implications of both early and late feature fusion. We demonstrate that both these approaches are suboptimal for a subset of object categories. Consequently, we propose a novel method for recognizing object categories when using multiple cues by separately processing the shape and color cues and combining them by modulating the shape features by category specific color attention. Color is used to compute bottom-up and top-down attention maps. Subsequently, the color attention maps are used to modulate the weights of the shape features. Shape features are given more weight in regions with higher attention and vice versa. The approach is tested on several benchmark object recognition data sets and the results clearly demonstrate the effectiveness of our proposed method. In the second part of the thesis, we investigate the problem of obtaining compact spatial pyramid representations for object and scene recognition. Spatial pyramids have been successfully applied to incorporate spatial information into bag-of-words based image representation. However, a major drawback of spatial pyramids is that it leads to high dimensional image representations. We present a novel framework for obtaining compact pyramid representation. The approach reduces the size of a high dimensional pyramid representation upto an order of magnitude without any significant reduction in accuracy. Moreover, we also investigate the optimal combination of multiple features such as color and shape within the context of our compact pyramid representation. Finally, we describe a novel technique to build discriminative visual words from multiple cues learned independently from training images. To this end, we use an information theoretic vocabulary compression technique to find discriminative combinations of visual cues and the resulting visual vocabulary is compact, has the cue binding property, and supports individual weighting of cues in the final image representation. The approach is tested on standard object recognition data sets. The results obtained clearly demonstrate the effectiveness of our approach.

Abstract: This thesis embarks on an exploratory journey into robustness in deep learning,
with a keen focus on the intertwining facets of generalization, explainability, and
edge cases within the realm of computer vision. In deep learning, robustness
epitomizes a model’s resilience and flexibility, grounded on its capacity to generalize across diverse data distributions, explain its predictions transparently, and navigate the intricacies of edge cases effectively. The challenges associated with robust generalization are multifaceted, encompassing the model’s performance on unseen data and its defense against out-of-distribution data and adversarial attacks. Bridging this gap, the potential of Embedding Propagation (EP) for improving out-of-distribution generalization is explored. EP is depicted as a powerful tool facilitating manifold smoothing, which in turn fortifies the model’s robustness against adversarial onslaughts and bolsters performance in few-shot and self-/semi-supervised learning scenarios. In the labyrinth of deep learning models, the path to robustness often intersects with explainability. As model complexity increases, so does the urgency to decipher their decision-making
processes. Acknowledging this, the thesis introduces a robust framework for
evaluating and comparing various counterfactual explanation methods, echoing
the imperative of explanation quality over quantity and spotlighting the intricacies of diversifying explanations. Simultaneously, the deep learning landscape is fraught with edge cases – anomalies in the form of small objects or rare instances in object detection tasks that defy the norm. Confronting this, the
thesis presents an extension of the DETR (DEtection TRansformer) model to enhance small object detection. The devised DETR-FP, embedding the Feature Pyramid technique, demonstrating improvement in small objects detection accuracy, albeit facing challenges like high computational costs. With emergence of foundation models in mind, the thesis unveils EarthView, the largest scale remote sensing dataset to date, built for the self-supervised learning of a robust foundational model for remote sensing. Collectively, these studies contribute to the grand narrative of robustness in deep learning, weaving together the strands of generalization, explainability, and edge case performance. Through these methodological advancements and novel datasets, the thesis calls for continued exploration, innovation, and refinement to fortify the bastion of robust computer vision.

Abstract: Unsupervised representation learning aims at finding methods that learn representations from data without annotation-based signals. Abstaining from annotations not only leads to economic benefits but may – and to some extent already does – result in advantages regarding the representation’s structure, robustness, and generalizability to different tasks. In the long run, unsupervised methods are expected to surpass their supervised counterparts due to the reduction of human intervention and the inherently more general setup that does not bias the optimization towards an objective originating from specific annotation-based signals. While major advantages of unsupervised representation learning have been recently observed in natural language processing, supervised methods still dominate in vision domains for most tasks. In this dissertation, we contribute to the field of unsupervised (visual) representation learning from three perspectives: (i) Learning representations: We design unsupervised, backpropagation-free Convolutional Self-Organizing Neural Networks (CSNNs) that utilize self-organization- and Hebbian-based learning rules to learn convolutional kernels and masks to achieve deeper backpropagation-free models. Thereby, we observe that backpropagation-based and -free methods can suffer from an objective function mismatch between the unsupervised pretext task and the target task. This mismatch can lead to performance decreases for the target task. (ii) Evaluating representations: We build upon the widely used (non-)linear evaluation protocol to define pretext- and target-objective-independent metrics for measuring the objective function mismatch. With these metrics, we evaluate various pretext and target tasks and disclose dependencies of the objective function mismatch concerning different parts of the training and model setup. (iii) Transferring representations: We contribute CARLANE, the first 3-way sim-to-real domain adaptation benchmark for 2D lane detection. We adopt several well-known unsupervised domain adaptation methods as baselines and propose a method based on prototypical cross-domain self-supervised learning. Finally, we focus on pixel-based unsupervised domain adaptation and contribute a content-consistent unpaired image-to-image translation method that utilizes masks, global and local discriminators, and similarity sampling to mitigate content inconsistencies, as well as feature-attentive denormalization to fuse content-based statistics into the generator stream. In addition, we propose the cKVD metric to incorporate class-specific content inconsistencies into perceptual metrics for measuring translation quality.

Abstract: The comprehension of animal behavior, especially human behavior, is one of the most ancient and studied problems since the beginning of civilization. The big list of factors that interact to determine a person action require the collaboration of different disciplines, such as psichology, biology, or sociology. In the last years the analysis of human behavior has received great attention also from the computer vision community, given the latest advances in the acquisition of human motion data from image sequences.
Despite the increasing availability of that data, there still exists a gap towards obtaining a conceptual representation of the obtained observations. Human behavior analysis is based on a qualitative interpretation of the results, and therefore the assignment of concepts to quantitative data is linked to a certain ambiguity.
This Thesis tackles the problem of obtaining a proper representation of human behavior in the contexts of computer vision and animation. On the one hand, a good behavior model should permit the recognition and explanation the observed activity in image sequences. On the other hand, such a model must allow the generation of new synthetic instances, which model the behavior of virtual agents.
First, we propose methods to automatically learn the models from observations. Given a set of quantitative results output by a vision system, a normal behavior model is learnt. This results provides a tool to determine the normality or abnormality of future observations. However, machine learning methods are unable to provide a richer description of the observations. We confront this problem by means of a new method that incorporates prior knowledge about the enviornment and about the expected behaviors. This framework, formed by the reasoning engine FMTL and the modeling tool SGT allows the generation of conceptual descriptions of activity in new image sequences. Finally, we demonstrate the suitability of the proposed framework to simulate behavior of virtual agents, which are introduced into real image sequences and interact with observed real agents, thereby easing the generation of augmented reality sequences.
The set of approaches presented in this Thesis has a growing set of potential applications. The analysis and description of behavior in image sequences has its principal application in the domain of smart video--surveillance, in order to detect suspicious or dangerous behaviors. Other applications include automatic sport commentaries, elderly monitoring, road traffic analysis, and the development of semantic video search engines. Alternatively, behavioral virtual agents allow to simulate accurate real situations, such as fires or crowds. Moreover, the inclusion of virtual agents into real image sequences has been widely deployed in the games and cinema industries.

Abstract: This Thesis is mainly divided in two parts. The first one presents a study of motion
segmentation problems. Based on this study, a novel algorithm for mobile-object
segmentation from a static background scene is also presented. This approach is
demonstrated robust and accurate under most of the common problems in motion
segmentation. The second one tackles the problem of shadows in depth. Firstly, a
bottom-up approach based on a chromatic shadow detector is presented to deal with
umbra shadows. Secondly, a top-down approach based on a tracking system has been
developed in order to enhance the chromatic shadow detection.
In our first contribution, a case analysis of motion segmentation problems is presented by taking into account the problems associated with different cues, namely
colour, edge and intensity. Our second contribution is a hybrid architecture which
handles the main problems observed in such a case analysis, by fusing (i) the knowledge from these three cues and (ii) a temporal difference algorithm. On the one hand,
we enhance the colour and edge models to solve both global/local illumination changes
(shadows and highlights) and camouflage in intensity. In addition, local information is
exploited to cope with a very challenging problem such as the camouflage in chroma.
On the other hand, the intensity cue is also applied when colour and edge cues are not
available, such as when beyond the dynamic range. Additionally, temporal difference
is included to segment motion when these three cues are not available, such as that
background not visible during the training period. Lastly, the approach is enhanced
for allowing ghost detection. As a result, our approach obtains very accurate and robust motion segmentation in both indoor and outdoor scenarios, as quantitatively and
qualitatively demonstrated in the experimental results, by comparing our approach
with most best-known state-of-the-art approaches.
Motion Segmentation has to deal with shadows to avoid distortions when detecting
moving objects. Most segmentation approaches dealing with shadow detection are
typically restricted to penumbra shadows. Therefore, such techniques cannot cope
well with umbra shadows. Consequently, umbra shadows are usually detected as part
of moving objects.
Firstly, a bottom-up approach for detection and removal of chromatic moving
shadows in surveillance scenarios is proposed. Secondly, a top-down approach based
on kalman filters to detect and track shadows has been developed in order to enhance
the chromatic shadow detection. In the Bottom-up part, the shadow detection approach applies a novel technique based on gradient and colour models for separating
chromatic moving shadows from moving objects.
Well-known colour and gradient models are extended and improved into an invariant colour cone model and an invariant gradient model, respectively, to perform
automatic segmentation while detecting potential shadows. Hereafter, the regions corresponding to potential shadows are grouped by considering ”a bluish effect” and an
edge partitioning. Lastly, (i) temporal similarities between local gradient structures
and (ii) spatial similarities between chrominance angle and brightness distortions are
analysed for all potential shadow regions in order to finally identify umbra shadows.
In the top-down process, after detection of objects and shadows both are tracked
using Kalman filters, in order to enhance the chromatic shadow detection, when it
fails to detect a shadow. Firstly, this implies a data association between the blobs
(foreground and shadow) and Kalman filters. Secondly, an event analysis of the different data association cases is performed, and occlusion handling is managed by a
Probabilistic Appearance Model (PAM). Based on this association, temporal consistency is looked for the association between foregrounds and shadows and their
respective Kalman Filters. From this association several cases are studied, as a result
lost chromatic shadows are correctly detected. Finally, the tracking results are used
as feedback to improve the shadow and object detection.
Unlike other approaches, our method does not make any a-priori assumptions
about camera location, surface geometries, surface textures, shapes and types of
shadows, objects, and background. Experimental results show the performance and
accuracy of our approach in different shadowed materials and illumination conditions.

Abstract: Psychological evidence has emphasized the importance of affective behaviour understanding due to its high impact in nowadays interaction humans and computers. All
type of affective and behavioural patterns such as gestures, emotions and mental
states are highly displayed through the face, head and body. Therefore, this thesis is
focused to analyse affective behaviours on head and face. To this end, head and facial
movements are encoded by using appearance based tracking methods. Specifically,
a wise combination of deformable models captures rigid and non-rigid movements of
different kinematics; 3D head pose, eyebrows, mouth, eyelids and irises are taken into
account as basis for extracting features from databases of video sequences. This approach combines the strengths of adaptive appearance models, optimization methods
and backtracking techniques.
For about thirty years, computer sciences have addressed the investigation on
human emotions to the automatic recognition of six prototypic emotions suggested
by Darwin and systematized by Paul Ekman in the seventies. The Facial Action
Coding System (FACS) which uses discrete movements of the face (called Action
units or AUs) to code the six facial emotions named anger, disgust, fear, happy-Joy,
sadness and surprise. However, human emotions are much complex patterns that
have not received the same attention from computer scientists.
Simon Baron-Cohen proposed a new taxonomy of emotions and mental states
without a system coding of the facial actions. These 426 affective behaviours are
more challenging for the understanding of human emotions. Beyond of classically
classifying the six basic facial expressions, more subtle gestures, facial actions and
spontaneous emotions are considered here. By assessing confidence on the recognition
results, exploring spatial and temporal relationships of the features, some methods are
combined and enhanced for developing new taxonomy of expressions and emotions.
The objective of this dissertation is to develop a computer vision system, including both facial feature extraction, expression recognition and emotion understanding
by building a bottom-up reasoning process. Building a detailed taxonomy of human
affective behaviours is an interesting challenge for head-face-based image analysis
methods. In this paper, we exploit the strengths of Canonical Correlation Analysis
(CCA) to enhance an on-line head-face tracker. A relationship between head pose and
local facial movements is studied according to their cognitive interpretation on affective expressions and emotions. Active Shape Models are synthesized for AAMs based
on CCA-regression. Head pose and facial actions are fused into a maximally correlated space in order to assess expressiveness, confidence and classification in a CBR system. The CBR solutions are also correlated to the cognitive features, which allow
avoiding exhaustive search when recognizing new head-face features. Subsequently,
Support Vector Machines (SVMs) and Bayesian Networks are applied for learning the
spatial relationships of facial expressions. Similarly, the temporal evolution of facial
expressions, emotion and mental states are analysed based on Factorized Dynamic
Bayesian Networks (FaDBN).
As results, the bottom-up system recognizes six facial expressions, six basic emotions and six mental states, plus enhancing this categorization with confidence assessment at each level, intensity of expressions and a complete taxonomy

Abstract: Humans can effortlessly perceive faces, follow them over space and time, and decode their rich content, such as pose, identity and expression. However, despite many decades of research on automatic facial perception in areas like face detection, expression recognition, pose estimation and face recognition, and despite many successes, a complete solution remains elusive. This thesis is dedicated to three problems in automatic face perception, namely face detection, face tracking and pose estimation.
In face detection, an initial simple model is presented that uses pixel-based heuristics to segment skin locations and hand-crafted rules to determine the locations of the faces present in an image. Different colorspaces are studied to judge whether a colorspace transformation can aid skin color detection. The output of this study is used in the design of a more complex face detector that is able to successfully generalize to different scenarios.
In face tracking, a framework that combines estimation and control in a joint scheme is presented to track a face with a single pan-tilt-zoom camera. While this work is mainly motivated by tracking faces, it can be easily applied atop of any detector to track different objects. The applicability of this method is demonstrated on simulated as well as real-life scenarios.
The last and most important part of this thesis is dedicate to monocular head pose estimation. In this part, a method based on partial least squares (PLS) regression is proposed to estimate pose and solve the alignment problem simultaneously. The contributions of this work are two-fold: 1) demonstrating that the proposed method achieves better than state-of-the-art results on the estimation problem and 2) developing a technique to reduce misalignment based on the learned PLS factors that outperform multiple instance learning (MIL) without the need for any re-training or the inclusion of misaligned samples in the training process, as normally done in MIL.