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: Writing is one of the most important forms of communication and for centuries, handwriting had been the most reliable way to preserve knowledge. However, despite the recent development of printing houses and electronic devices, handwriting is still broadly used for taking notes, doing annotations, or sketching ideas.
Transferring the ability of understanding handwritten text or recognizing handwritten shapes to computers has been the goal of many researches due to its huge importance for many different fields. However, designing good representations to deal with handwritten shapes, e.g. symbols or words, is a very challenging problem due to the large variability of these kinds of shapes. One of the consequences of working with handwritten shapes is that we need representations to be robust, i.e., able to adapt to large intra-class variability. We need representations to be discriminative, i.e., able to learn what are the differences between classes. And, we need representations to be efficient, i.e., able to be rapidly computed and compared. Unfortunately, current techniques of handwritten shape representation for matching and recognition do not fulfill some or all of these requirements.
Through this thesis we focus on the problem of learning to represent handwritten shapes aimed at retrieval and recognition tasks. Concretely, on the first part of the thesis, we focus on the general problem of representing any kind of handwritten shape. We first present a novel shape descriptor based on a deformable grid that deals with large deformations by adapting to the shape and where the cells of the grid can be used to extract different features. Then, we propose to use this descriptor to learn statistical models, based on the Active Appearance Model, that jointly learns the variability in structure and texture of a given class. Then, on the second part, we focus on a concrete application, the problem of representing handwritten words, for the tasks of word spotting, where the goal is to find all instances of a query word in a dataset of images, and recognition. First, we address the segmentation-free problem and propose an unsupervised, sliding-window-based approach that achieves state-of- the-art results in two public datasets. Second, we address the more challenging multi-writer problem, where the variability in words exponentially increases. We describe an approach in which both word images and text strings are embedded in a common vectorial subspace, and where those that represent the same word are close together. This is achieved by a combination of label embedding and attributes learning, and a common subspace regression. This leads to a low-dimensional, unified representation of word images and strings, resulting in a method that allows one to perform either image and text searches, as well as image transcription, in a unified framework. We evaluate our methods on different public datasets of both handwritten documents and natural images showing results comparable or better than the state-of-the-art on spotting and recognition tasks.

Abstract: In this thesis we explore information Extraction from totally or partially handwritten documents. Basically we are dealing with two different application scenarios. The first scenario are modern highly structured documents like forms. In this kind of documents, the semantic information is encoded in different fields with a pre-defined location in the document, therefore, information extraction becomes roughly equivalent to transcription. The second application scenario are loosely structured totally handwritten documents, besides transcribing them, we need to assign a semantic label, from a set of known values to the handwritten words.
In both scenarios, transcription is an important part of the information extraction. For that reason in this thesis we present two methods based on Neural Networks, to transcribe handwritten text.In order to tackle the challenge of loosely structured documents, we have produced a benchmark, consisting of a dataset, a defined set of tasks and a metric, that was presented to the community as an international competition. Also, we propose different models based on Convolutional and Recurrent neural networks that are able to transcribe and assign different semantic labels to each handwritten words, that is, able to perform Information Extraction.

Abstract: More realistic and accurate models of the scene illumination and object reflectance can greatly improve the quality of many computer vision and computer graphics tasks. Using such model, a more profound knowledge about the interaction of light with object surfaces can be established which proves crucial to a variety of computer vision applications. In the current work, we investigate the various existing approaches to illumination and reflectance modeling and form an analysis on their shortcomings in capturing the complexity of real-world scenes. Based on this analysis we propose improvements to different aspects of reflectance and illumination estimation in order to more realistically model the real-world scenes in the presence of complex lighting phenomena (i.e, multiple illuminants, interreflections and shadows). Moreover, we captured our own multi-illuminant dataset which consists of complex scenes and illumination conditions both outdoor and in laboratory conditions. In addition we investigate the use of synthetic data to facilitate the construction of datasets and improve the process of obtaining ground-truth information.

Abstract: Automatic analyzing humans in photographs or videos has great potential applications in computer vision, including medical diagnosis, sports, entertainment, movie editing and surveillance, just to name a few. Body, face and hand are the most studied components of humans. Body has many variabilities in shape and clothing along with high degrees of freedom in pose. Face has many muscles causing many visible deformity, beside variable shape and hair style. Hand is a small object, moving fast and has high degrees of freedom. Adding human characteristics to all aforementioned variabilities makes human analysis quite a challenging task.
In this thesis, we developed human segmentation in different modalities. In a first scenario, we segmented human body and hand in depth images using example-based shape warping. We developed a shape descriptor based on shape context and class probabilities of shape regions to extract nearest neighbors. We then considered rigid affine alignment vs. nonrigid iterative shape warping. In a second scenario, we segmented face in RGB images using convolutional neural networks (CNN). We modeled conditional random field with recurrent neural networks. In our model pair-wise kernels are not fixed and learned during training. We trained the network end-to-end using adversarial networks which improved hair segmentation by a high margin.
We also worked on 3D hand pose estimation in depth images. In a generative approach, we fitted a finger model separately for each finger based on our example-based rigid hand segmentation. We minimized an energy function based on overlapping area, depth discrepancy and finger collisions. We also applied linear models in joint trajectory space to refine occluded joints based on visible joints error and invisible joints trajectory smoothness. In a CNN-based approach, we developed a tree-structure network to train specific features for each finger and fused them for global pose consistency. We also formulated physical and appearance constraints as loss functions.
Finally, we developed a number of applications consisting of human soft biometrics measurement and garment retexturing. We also generated some datasets in this thesis consisting of human segmentation, synthetic hand pose, garment retexturing and Italian gestures.