Abstract: Video alignment consists of integrating multiple video sequences recorded independently into a single video sequence. This means to register both in time (synchronize
frames) and space (image registration) so that the two videos sequences can be fused
or compared pixel–wise. In spite of being relatively unknown, many applications today may benefit from the availability of robust and efficient video alignment methods.
For instance, video surveillance requires to integrate video sequences that are recorded
of the same scene at different times in order to detect changes. The problem of aligning videos has been addressed before, but in the relatively simple cases of fixed or rigidly attached cameras and simultaneous acquisition. In addition, most works rely
on restrictive assumptions which reduce its difficulty such as linear time correspondence or the knowledge of the complete trajectories of corresponding scene points on the images; to some extent, these assumptions limit the practical applicability of the solutions developed until now. In this thesis, we focus on the challenging problem of aligning sequences recorded at different times from independent moving cameras following similar but not coincident trajectories. More precisely, this thesis covers four studies that advance the state-of-the-art in video alignment. First, we focus on analyzing and developing a probabilistic framework for video alignment, that is, a principled way to integrate multiple observations and prior information. In this way, two different approaches are presented to exploit the combination of several purely visual features (image–intensities, visual words and dense motion field descriptor), and
global positioning system (GPS) information. Second, we focus on reformulating the
problem into a single alignment framework since previous works on video alignment
adopt a divide–and–conquer strategy, i.e., first solve the synchronization, and then
register corresponding frames. This also generalizes the ’classic’ case of fixed geometric transform and linear time mapping. Third, we focus on exploiting directly the
time domain of the video sequences in order to avoid exhaustive cross–frame search.
This provides relevant information used for learning the temporal mapping between
pairs of video sequences. Finally, we focus on adapting these methods to the on–line
setting for road detection and vehicle geolocation. The qualitative and quantitative
results presented in this thesis on a variety of real–world pairs of video sequences show that the proposed method is: robust to varying imaging conditions, different image
content (e.g., incoming and outgoing vehicles), variations on camera velocity, and
different scenarios (indoor and outdoor) going beyond the state–of–the–art. Moreover, the on–line video alignment has been successfully applied for road detection and
vehicle geolocation achieving promising results.

Abstract: Document Visual Question Answering (DocVQA) is a fast growing branch of document understanding. Despite the fact that documents contain sensitive or copyrighted information, none of the current DocVQA methods offers strong privacy guarantees.
In this work, we explore privacy in the domain of DocVQA for the first time. We highlight privacy issues in state of the art multi-modal LLM models used for DocVQA, and explore possible solutions.
Specifically, we focus on the invoice processing use case as a realistic, widely used scenario for document understanding, and propose a large scale DocVQA dataset comprising invoice documents and associated questions and answers. We employ a federated learning scheme, that reflects the real-life distribution of documents in different businesses, and we explore the use case where the ID of the invoice issuer is the sensitive information to be protected.
We demonstrate that non-private models tend to memorise, behaviour that can lead to exposing private information. We then evaluate baseline training schemes employing federated learning and differential privacy in this multi-modal scenario, where the sensitive information might be exposed through any of the two input modalities: vision (document image) or language (OCR tokens).
Finally, we design an attack exploiting the memorisation effect of the model, and demonstrate its effectiveness in probing different DocVQA models.

Abstract: In Part 1 of the thesis, we hypothesize that salient and non-salient image regions can be estimated to be the regions which are enhanced or assimilated in standard low-level color image representations. We prove this hypothesis by adapting a low-level model of color perception into a saliency estimation model. This model shares the three main steps found in many successful models for predicting attention in a scene: convolution with a set of filters, a center-surround mechanism and spatial pooling to construct a saliency map. For such models, integrating spatial information and justifying the choice of various parameter values remain open problems. Our saliency model inherits a principled selection of parameters as well as an innate spatial pooling mechanism from the perception model on which it is based. This pooling mechanism has been fitted using psychophysical data acquired in color-luminance setting experiments. The proposed model outperforms the state-of-the-art at the task of predicting eye-fixations from two datasets. After demonstrating the effectiveness of our basic saliency model, we introduce an improved image representation, based on geometrical grouplets, that enhances complex low-level visual features such as corners and terminations, and suppresses relatively simpler features such as edges. With this improved image representation, the performance of our saliency model in predicting eye-fixations increases for both datasets.
In Part 2 of the thesis, we investigate the problem of aesthetic visual analysis. While a great deal of research has been conducted on hand-crafting image descriptors for aesthetics, little attention so far has been dedicated to the collection, annotation and distribution of ground truth data. Because image aesthetics is complex and subjective, existing datasets, which have few images and few annotations, have significant limitations. To address these limitations, we have introduced a new large-scale database for conducting Aesthetic Visual Analysis, which we call AVA. AVA contains more than 250,000 images, along with a rich variety of annotations. We investigate how the wealth of data in AVA can be used to tackle the challenge of understanding and assessing visual aesthetics by looking into several problems relevant for aesthetic analysis. We demonstrate that by leveraging the data in AVA, and using generic low-level features such as SIFT and color histograms, we can exceed state-of-the-art performance in aesthetic quality prediction tasks.
Finally, we entertain the hypothesis that low-level visual information in our saliency model can also be used to predict visual aesthetics by capturing local image characteristics such as feature contrast, grouping and isolation, characteristics thought to be related to universal aesthetic laws. We use the weighted center-surround responses that form the basis of our saliency model to create a feature vector that describes aesthetics. We also introduce a novel color space for fine-grained color representation. We then demonstrate that the resultant features achieve state-of-the-art performance on aesthetic quality classification.
As such, a promising contribution of this thesis is to show that several vision experiences – low-level color perception, visual saliency and visual aesthetics estimation – may be successfully modeled using a unified framework. This suggests a similar architecture in area V1 for both color perception and saliency and adds evidence to the hypothesis that visual aesthetics appreciation is driven in part by low-level cues.

Abstract: The amount of multimodal devices that surround us is growing everyday. In this context, human interaction and communication have become a focus of attention and a hot topic of research. A crucial element in human relations is the evaluation of individuals with respect to facial traits, what is called a first impression. Studies based on appearance have suggested that personality can be expressed by appearance and the observer may use such information to form judgments. In the context of rapid facial evaluation, certain personality traits seem to have a more pronounced effect on the relations and perceptions inside groups. The perception of dominance has been shown to be an active part of social roles at different stages of life, and even play a part in mate selection. The aim of this paper is to study to what extent this information is learnable from the point of view of computer science. Specifically we intend to determine if judgments of dominance can be learned by machine learning techniques. We implement two different descriptors in order to assess this. The first is the histogram of oriented gradients (HOG), and the second is a probabilistic appearance descriptor based on the frequencies of grouped binary tests. State of the art classification rules validate the performance of both descriptors, with respect to the prediction task. Experimental results show that machine learning techniques can predict judgments of dominance rather accurately (accuracies up to 90%) and that the HOG descriptor may characterize appropriately the information necessary for such task.