Abstract: Recognizing visual categories from semantic descriptions is a promising way to extend the capability of a visual classifier beyond the concepts represented in the training data (i.e. seen categories). This problem is addressed by (generalized) zero-shot learning methods (GZSL), which leverage semantic descriptions that connect them to seen categories (e.g. label embedding, attributes). Conventional GZSL are designed mostly for object recognition. In this paper we focus on zero-shot scene recognition, a more challenging setting with hundreds of categories where their differences can be subtle and often localized in certain objects or regions. Conventional GZSL representations are not rich enough to capture these local discriminative differences. Addressing these limitations, we propose a feature generation framework with two novel components: 1) multiple sources of semantic information (i.e. attributes, word embeddings and descriptions), 2) region descriptions that can enhance scene discrimination. To generate synthetic visual features we propose a two-step generative approach, where local descriptions are sampled and used as conditions to generate visual features. The generated features are then aggregated and used together with real features to train a joint classifier. In order to evaluate the proposed method, we introduce a new dataset for zero-shot scene recognition with multi-semantic annotations. Experimental results on the proposed dataset and SUN Attribute dataset illustrate the effectiveness of the proposed method.

Abstract: For humans, the recognition of objects is an almost instantaneous, precise and
extremely adaptable process. Furthermore, we have the innate capability to learn
new object classes from only few examples. The human brain lowers the complexity
of the incoming data by filtering out part of the information and only processing
those things that capture our attention. This, mixed with our biological predisposition to respond to certain shapes or colors, allows us to recognize in a simple
glance the most important or salient regions from an image. This mechanism can
be observed by analyzing on which parts of images subjects place attention; where
they fix their eyes when an image is shown to them. The most accurate way to
record this behavior is to track eye movements while displaying images.
Computational saliency estimation aims to identify to what extent regions or
objects stand out with respect to their surroundings to human observers. Saliency
maps can be used in a wide range of applications including object detection, image
and video compression, and visual tracking. The majority of research in the field has
focused on automatically estimating saliency maps given an input image. Instead, in
this thesis, we set out to incorporate saliency maps in an object recognition pipeline:
we want to investigate whether saliency maps can improve object recognition
results.
In this thesis, we identify several problems related to visual saliency estimation.
First, to what extent the estimation of saliency can be exploited to improve the
training of an object recognition model when scarce training data is available. To
solve this problem, we design an image classification network that incorporates
saliency information as input. This network processes the saliency map through a
dedicated network branch and uses the resulting characteristics to modulate the
standard bottom-up visual characteristics of the original image input. We will refer to this technique as saliency-modulated image classification (SMIC). In extensive
experiments on standard benchmark datasets for fine-grained object recognition,
we show that our proposed architecture can significantly improve performance,
especially on dataset with scarce training data.
Next, we address the main drawback of the above pipeline: SMIC requires an
explicit saliency algorithm that must be trained on a saliency dataset. To solve this,
we implement a hallucination mechanism that allows us to incorporate the saliency
estimation branch in an end-to-end trained neural network architecture that only
needs the RGB image as an input. A side-effect of this architecture is the estimation
of saliency maps. In experiments, we show that this architecture can obtain similar
results on object recognition as SMIC but without the requirement of ground truth
saliency maps to train the system.
Finally, we evaluated the accuracy of the saliency maps that occur as a sideeffect of object recognition. For this purpose, we use a set of benchmark datasets
for saliency evaluation based on eye-tracking experiments. Surprisingly, the estimated saliency maps are very similar to the maps that are computed from human
eye-tracking experiments. Our results show that these saliency maps can obtain
competitive results on benchmark saliency maps. On one synthetic saliency dataset
this method even obtains the state-of-the-art without the need of ever having seen
an actual saliency image for training.

Abstract: Outlier, or anomaly, detection is essential for optimal performance of machine learning methods and statistical predictive models. It is not just a technical step in a data cleaning process but a key topic in many fields such as fraudulent document detection, in medical applications and assisted diagnosis systems or detecting security threats. In contrast to population-based methods, neighborhood based local approaches are simple flexible methods that have the potential to perform well in small sample size unbalanced problems. However, a main concern of local approaches is the impact that the computation of each sample neighborhood has on the method performance. Most approaches use a distance in the feature space to define a single neighborhood that requires careful selection of several parameters. This work presents a local approach based on a local measure of the heterogeneity of sample labels in the feature space considered as a topological manifold. Topology is computed using the communities of a weighted graph codifying mutual nearest neighbors in the feature space. This way, we provide with a set of multiple neighborhoods able to describe the structure of complex spaces without parameter fine tuning. The extensive experiments on real-world data sets show that our approach overall outperforms, both, local and global strategies in multi and single view settings.

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: Machine learning experimentation under controlled scenarios and standard datasets is necessary to compare algorithms performance by evaluating all of them in the same setup. However, experimentation on how those algorithms perform on unconstrained data and applied tasks to solve real world problems is also a must to ascertain how that research can contribute to our society.
In this dissertation we experiment with the latest computer vision and natural language processing algorithms applying them to multimodal scene interpretation. Particularly, we research on how image and text understanding can be jointly exploited to address real world problems, focusing on learning from Social Media data.
We address several tasks that involve image and textual information, discuss their characteristics and offer our experimentation conclusions. First, we work on detection of scene text in images. Then, we work with Social Media posts, exploiting the captions associated to images as supervision to learn visual features, which we apply to multimodal semantic image retrieval. Subsequently, we work with geolocated Social Media images with associated tags, experimenting on how to use the tags as supervision, on location sensitive image retrieval and on exploiting location information for image tagging. Finally, we work on a specific classification problem of Social Media publications consisting on an image and a text: Multimodal hate speech classification.

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.