Abstract: Manually annotating images to develop vision models has been a major bottleneck
since computer vision and machine learning started to walk together. This has
been more evident since computer vision falls on the shoulders of data-hungry
deep learning techniques. When addressing on-board perception for autonomous
driving, the curse of data annotation is exacerbated due to the use of additional
sensors such as LiDAR. Therefore, any approach aiming at reducing such a timeconsuming and costly work is of high interest for addressing autonomous driving
and, in fact, for any application requiring some sort of artificial perception. In the
last decade, it has been shown that leveraging from synthetic data is a paradigm
worth to pursue in order to minimizing manual data annotation. The reason is
that the automatic process of generating synthetic data can also produce different
types of associated annotations (e.g. object bounding boxes for synthetic images
and LiDAR pointclouds, pixel/point-wise semantic information, etc.). Directly
using synthetic data for training deep perception models may not be the definitive
solution in all circumstances since it can appear a synth-to-real domain shift. In
this context, this work focuses on leveraging synthetic data to alleviate manual
annotation for three perception tasks related to driving assistance and autonomous
driving. In all cases, we assume the use of deep convolutional neural networks
(CNNs) to develop our perception models.
The first task addresses traffic sign recognition (TSR), a kind of multi-class
classification problem. We assume that the number of sign classes to be recognized
must be suddenly increased without having annotated samples to perform the
corresponding TSR CNN re-training. We show that leveraging synthetic samples of
such new classes and transforming them by a generative adversarial network (GAN)
trained on the known classes (i.e. without using samples from the new classes), it is
possible to re-train the TSR CNN to properly classify all the signs for a ∼ 1/4 ratio of
new/known sign classes. The second task addresses on-board 2D object detection,
focusing on vehicles and pedestrians. In this case, we assume that we receive a set
of images without the annotations required to train an object detector, i.e. without
object bounding boxes. Therefore, our goal is to self-annotate these images so
that they can later be used to train the desired object detector. In order to reach
this goal, we leverage from synthetic data and propose a semi-supervised learning
approach based on the co-training idea. In fact, we use a GAN to reduce the synthto-real domain shift before applying co-training. Our quantitative results show
that co-training and GAN-based image-to-image translation complement each
other up to allow the training of object detectors without manual annotation, and still almost reaching the upper-bound performances of the detectors trained from
human annotations. While in previous tasks we focus on vision-based perception,
the third task we address focuses on LiDAR pointclouds. Our initial goal was to
develop a 3D object detector trained on synthetic LiDAR-style pointclouds. While
for images we may expect synth/real-to-real domain shift due to differences in
their appearance (e.g. when source and target images come from different camera
sensors), we did not expect so for LiDAR pointclouds since these active sensors
factor out appearance and provide sampled shapes. However, in practice, we have
seen that it can be domain shift even among real-world LiDAR pointclouds. Factors
such as the sampling parameters of the LiDARs, the sensor suite configuration onboard the ego-vehicle, and the human annotation of 3D bounding boxes, do induce
a domain shift. We show it through comprehensive experiments with different
publicly available datasets and 3D detectors. This redirected our goal towards the
design of a GAN for pointcloud-to-pointcloud translation, a relatively unexplored
topic.
Finally, it is worth to mention that all the synthetic datasets used for these three
tasks, have been designed and generated in the context of this PhD work and will
be publicly released. Overall, we think this PhD presents several steps forward to
encourage leveraging synthetic data for developing deep perception models in the
field of driving assistance and autonomous driving.

Abstract: Computer vision has gone through considerable changes in the last decade as neural networks have come into common use. As available computational capabilities have grown, neural networks have achieved breakthroughs in many computer vision tasks, and have even surpassed human performance in others. With accuracy being so high, focus has shifted to other issues and challenges. One research direction that saw a notable increase in interest is on lifelong learning systems. Such systems should be capable of efficiently performing tasks, identifying and learning new ones, and should moreover be able to deploy smaller versions of themselves which are experts on specific tasks. In this thesis, we contribute to research on lifelong learning and address the compression and adaptation of networks to small target domains, the incremental learning of networks faced with a variety of tasks, and finally the detection of out-of-distribution samples at inference time.
We explore how knowledge can be transferred from large pretrained models to more task-specific networks capable of running on smaller devices by extracting the most relevant information. Using a pretrained model provides more robust representations and a more stable initialization when learning a smaller task, which leads to higher performance and is known as domain adaptation. However, those models are too large for certain applications that need to be deployed on devices with limited memory and computational capacity. In this thesis we show that, after performing domain adaptation, some learned activations barely contribute to the predictions of the model. Therefore, we propose to apply network compression based on low-rank matrix decomposition using the activation statistics. This results in a significant reduction of the model size and the computational cost.
Like human intelligence, machine intelligence aims to have the ability to learn and remember knowledge. However, when a trained neural network is presented with learning a new task, it ends up forgetting previous ones. This is known as catastrophic forgetting and its avoidance is studied in continual learning. The work presented in this thesis extensively surveys continual learning techniques and presents an approach to avoid catastrophic forgetting in sequential task learning scenarios. Our technique is based on using ternary masks in order to update a network to new tasks, reusing the knowledge of previous ones while not forgetting anything about them. In contrast to earlier work, our masks are applied to the activations of each layer instead of the weights. This considerably reduces the number of parameters to be added for each new task. Furthermore, the analysis on a wide range of work on incremental learning without access to the task-ID, provides insight on current state-of-the-art approaches that focus on avoiding catastrophic forgetting by using regularization, rehearsal of previous tasks from a small memory, or compensating the task-recency bias.
Neural networks trained with a cross-entropy loss force the outputs of the model to tend toward a one-hot encoded vector. This leads to models being too overly confident when presented with images or classes that were not present in the training distribution. The capacity of a system to be aware of the boundaries of the learned tasks and identify anomalies or classes which have not been learned yet is key to lifelong learning and autonomous systems. In this thesis, we present a metric learning approach to out-of-distribution detection that learns the task at hand on an embedding space.

Abstract: Over the last few years, the number of consumer computer vision applications has increased dramatically. Today, computer vision solutions can be found in video game consoles, smartphone applications, driving assistance – just to name a few. Ideally, we require the performance of those applications, particularly those that are safety critical to remain constant under any external environment factors, such as changes in illumination or weather conditions. However, this is not always possible or very difficult to obtain by only using visible imagery, due to the inherent limitations of the images from that spectral band. For that reason, the use of images from different or multiple spectral bands is becoming more appealing.
The aforementioned possible advantages of using images from multiples spectral bands on various vision applications make multi-spectral image processing a relevant topic for research and development. Like in visible image processing, multi-spectral image processing needs tools and algorithms to handle information from various spectral bands. Furthermore, traditional tools such as local feature detection, which is the basis of many vision tasks such as visual odometry, image registration, or structure from motion, must be adjusted or reformulated to operate under new conditions. Traditional feature detection, description, and matching methods tend to underperform in multi-spectral settings, in comparison to mono-spectral settings, due to the natural differences between each spectral band.
The work in this thesis is focused on the local feature description problem when cross-spectral images are considered. In this context, this dissertation has three main contributions. Firstly, the work starts by proposing the usage of a combination of frequency and spatial information, in a multi-scale scheme, as feature description. Evaluations of this proposal, based on classical hand-made feature descriptors, and comparisons with state of the art cross-spectral approaches help to find and understand limitations of such strategy. Secondly, different convolutional neural network (CNN) based architectures are evaluated when used to describe cross-spectral image patches. Results showed that CNN-based methods, designed to work with visible monocular images, could be successfully applied to the description of images from two different spectral bands, with just minor modifications. In this framework, a novel CNN-based network model, specifically intended to describe image patches from two different spectral bands, is proposed. This network, referred to as Q-Net, outperforms state of the art in the cross-spectral domain, including both previous hand-made solutions as well as L2 CNN-based architectures. The third contribution of this dissertation is in the cross-spectral feature description application domain. The multispectral odometry problem is tackled showing a real application of cross-spectral descriptors
In addition to the three main contributions mentioned above, in this dissertation, two different multi-spectral datasets are generated and shared with the community to be used as benchmarks for further studies.

Abstract: Feature matching is a fundamental problem in Computer Vision, having multiple applications such as tracking, image classification and retrieval, shape recognition and stereo fusion. In numerous domains, it is useful to represent the local structure of the matching features to increase the matching accuracy or to make the correspondence invariant to certain transformations (affine, homography, etc. . . ). However, encoding this knowledge requires complicating the model by establishing high-order relationships between the model elements, and therefore increasing the complexity of the optimization problem.
The importance of many-to-many matching is sometimes dismissed in the literature. Most methods are restricted to perform one-to-one matching, and are usually validated on synthetic, or non-realistic datasets. In a real challenging environment, with scale, pose and illumination variations of the object of interest, as well as the presence of occlusions, clutter, and noisy observations, many-to-many matching is necessary to achieve satisfactory results. As a consequence, finding the most likely many-to-many correspondence often involves a challenging combinatorial optimization process.
In this work, we design and demonstrate matching algorithms that compute many-to-many correspondences, applied to several challenging problems. Our goal is to make use of high-order representations to improve the expressive power of the matching, at the same time that we make feasible the process of inference or optimization of such models. We effectively use graphical models as our preferred representation because they provide an elegant probabilistic framework to tackle structured prediction problems.
We introduce a matching-based tracking algorithm which performs matching between frames of a video sequence in order to solve the difficult problem of headlight tracking at night-time. We also generalise this algorithm to solve the problem of data association applied to various tracking scenarios. We demonstrate the effectiveness of such approach in real video sequences and we show that our tracking algorithm can be used to improve the accuracy of a headlight classification system.
In the second part of this work, we move from single (point) matching to dense (region) matching and we introduce a new hierarchical image representation. We make use of such model to develop a high-order many-to-many matching between pairs of images. We show that the use of high-order models in comparison to simpler models improves not only the accuracy of the results, but also the convergence speed of the inference algorithm.
Finally, we keep exploiting the idea of region matching to design a fully unsupervised image co-segmentation algorithm that is able to perform competitively with state-of-the-art supervised methods. Our method also overcomes the typical drawbacks of some of the past works, such as avoiding the necessity of variate appearances on the image backgrounds. The region matching in this case is applied to effectively exploit inter-image information. We also extend this work to perform co-segmentation of videos, being the first time that such problem is addressed, as a way to perform video object segmentation

Abstract: Image values are the result of a combination of visual information coming from multiple sources. Recovering information from the multiple factors thatproduced an image seems a hard and ill-posed problem. However, it is important to observe that humans develop the ability to interpret images and recognize and isolate specific physical properties of the scene.
Images describing a single physical characteristic of an scene are called intrinsic images. These images would benefit most computer vision tasks which are often affected by the multiple complex effects that are usually found in natural images (e.g. cast shadows, specularities, interreflections...).
In this thesis we analyze the problem of intrinsic image estimation from different perspectives, including the theoretical formulation of the problem, the visual cues that can be used to estimate the intrinsic components and the evaluation mechanisms of the problem.