Abstract: One direct consequence of recent advances in digital visual data generation and the direct availability of this information through the World-Wide Web, is a urgent demand for efficient image retrieval systems. The objective of image retrieval is to allow users to efficiently browse through this abundance of images. Due to the non-expert nature of the majority of the internet users, such systems should be user friendly, and therefore avoid complex user interfaces. In this chapter we investigate how high-level information provided by recently developed object recognition techniques can improve interactive image retrieval. Wel apply a bagof- word based image representation method to automatically classify images in a number of categories. These additional labels are then applied to improve the image retrieval system. Next to these high-level semantic labels, we also apply a low-level image description to describe the composition and color scheme of the scene. Both descriptions are incorporated in a user feedback image retrieval setting. The main objective is to show that automatic labeling of images with semantic labels can improve image retrieval results.

Abstract: On-board pedestrian detection is crucial for Advanced Driver Assistance Systems
(ADAS). An accurate classication is fundamental for vision-based pedestrian detection.
The underlying assumption for learning classiers is that the training set and the deployment environment (testing) follow the same probability distribution regarding the features used by the classiers. However, in practice, there are dierent reasons that can break this constancy assumption. Accordingly, reusing existing classiers by adapting them from the previous training environment (source domain) to the new testing one (target domain) is an approach with increasing acceptance in the computer vision community. In this thesis we focus on the domain adaptation of deformable part-based models (DPMs) for pedestrian detection. As a prof of concept, we use a computer graphic based synthetic dataset, i.e. a virtual world, as the source domain, and adapt the virtual-world trained DPM detector to various real-world dataset.
We start by exploiting the maximum detection accuracy of the virtual-world
trained DPM. Even though, when operating in various real-world datasets, the virtualworld trained detector still suer from accuracy degradation due to the domain gap of virtual and real worlds. We then focus on domain adaptation of DPM. At the rst step, we consider single source and single target domain adaptation and propose two batch learning methods, namely A-SSVM and SA-SSVM. Later, we further consider leveraging multiple target (sub-)domains for progressive domain adaptation and propose a hierarchical adaptive structured SVM (HA-SSVM) for optimization. Finally, we extend HA-SSVM for the challenging online domain adaptation problem, aiming at making the detector to automatically adapt to the target domain online, without any human intervention. All of the proposed methods in this thesis do not require
revisiting source domain data. The evaluations are done on the Caltech pedestrian detection benchmark. Results show that SA-SSVM slightly outperforms A-SSVM and avoids accuracy drops as high as 15 points when comparing with a non-adapted detector. The hierarchical model learned by HA-SSVM further boosts the domain adaptation performance. Finally, the online domain adaptation method has demonstrated that it can achieve comparable accuracy to the batch learned models while not requiring manually label target domain examples. Domain adaptation for pedestrian detection is of paramount importance and a relatively unexplored area. We humbly hope the work in this thesis could provide foundations for future work in this area.

Abstract: The curse of data labeling is a costly bottleneck in supervised deep learning, where large amounts of labeled data are needed to train intelligent systems. In onboard perception for autonomous driving, this cost corresponds to the labeling of raw data from sensors such as cameras, LiDARs, RADARs, etc. Therefore, synthetic data with automatically generated ground truth (labels) has aroused as a reliable alternative for training onboard perception models.
However, synthetic data commonly suffers from synth-to-real domain shift, i.e., models trained on the synthetic domain do not show their achievable accuracy when performing in the real world. This shift needs to be addressed by techniques falling in the realm of domain adaptation (DA).
The semi-supervised learning (SSL) paradigm can be followed to address DA. In this case, a model is trained using source data with labels (here synthetic) and leverages minimal knowledge from target data (here the real world) to generate pseudo-labels. These pseudo-labels help the training process to reduce the gap between the source and the target domains. In general, we can assume accessing both, pseudo-labels and a few amounts of human-provided labels for the target-domain data. However, the most interesting and challenging setting consists in assuming that we do not have human-provided labels at all. This setting is known as unsupervised domain adaptation (UDA). This PhD focuses on applying SSL to the UDA setting, for onboard visual tasks related to autonomous driving. We start by addressing the synth-to-real UDA problem on onboard vision-based object detection (pedestrians and cars), a critical task for autonomous driving and driving assistance. In particular, we propose to apply an SSL technique known as co-training, which we adapt to work with deep models that process a multi-modal input. The multi-modality consists of the visual appearance of the images (RGB) and their monocular depth estimation. The synthetic data we use as the source domain contains both, object bounding boxes and depth information. This prior knowledge is the
starting point for the co-training technique, which iteratively labels unlabeled real-world data and uses such pseudolabels (here bounding boxes with an assigned object class) to progressively improve the labeling results. Along this
process, two models collaborate to automatically label the images, in a way that one model compensates for the errors of the other, so avoiding error drift. While this automatic labeling process is done offline, the resulting pseudolabels can be used to train object detection models that must perform in real-time onboard a vehicle. We show that multi-modal co-training improves the labeling results compared to single-modal co-training, remaining competitive compared to human labeling.
Given the success of co-training in the context of object detection, we have also adapted this technique to a more crucial and challenging visual task, namely, onboard semantic segmentation. In fact, providing labels for a single image
can take from 30 to 90 minutes for a human labeler, depending on the content of the image. Thus, developing automatic labeling techniques for this visual task is of great interest to the automotive industry. In particular, the new co-training framework addresses synth-to-real UDA by an initial stage of self-training. Intermediate models arising from this stage are used to start the co-training procedure, for which we have elaborated an accurate collaboration policy between the two models performing the automatic labeling. Moreover, our co-training seamlessly leverages datasets from different synthetic domains. In addition, the co-training procedure is agnostic to the loss function used to train the semantic segmentation models which perform the automatic labeling. We achieve state-of-the-art results on publicly available benchmark datasets, again, remaining competitive compared to human labeling.
Finally, on the ground of our previous experience, we have designed and implemented a new SSL technique for UDA in the context of visual semantic segmentation. In this case, we mimic the labeling methodology followed by human labelers. In particular, rather than labeling full images at a time, categories of semantic classes are defined and only those are labeled in a labeling pass. In fact, different human labelers can become specialists in labeling different categories. Afterward, these per-category-labeled layers are combined to provide fully labeled images. Our technique is inspired by this methodology since we perform synth-to-real UDA per category, using the self-training stage previously developed as part of our co-training framework. The pseudo-labels obtained for each category are finally
fused to obtain fully automatically labeled images. In this context, we have also contributed to the development of a new photo-realistic synthetic dataset based on path-tracing rendering. Our new SSL technique seamlessly leverages publicly available synthetic datasets as well as this new one to obtain state-of-the-art results on synth-to-real UDA for semantic segmentation. We show that the new dataset allows us to reach better labeling accuracy than previously existing datasets, at the same time that it complements well them when combined. Moreover, we also show that the new human-inspired SSL technique outperforms co-training.

Abstract: Pedestrian detection is of paramount interest for many applications, e.g. Advanced Driver Assistance Systems, Intelligent Video Surveillance and Multimedia systems. Most promising pedestrian detectors rely on appearance-based classifiers trained with annotated data. However, the required annotation step represents an intensive and subjective task for humans, what makes worth to minimize their intervention in this process by using computational tools like realistic virtual worlds. The reason to use these kind of tools relies in the fact that they allow the automatic generation of precise and rich annotations of visual information. Nevertheless, the use of this kind of data comes with the following question: can a pedestrian appearance model learnt with virtual-world data work successfully for pedestrian detection in real-world scenarios?. To answer this question, we conduct different experiments that suggest a positive answer. However, the pedestrian classifiers trained with virtual-world data can suffer the so called dataset shift problem as real-world based classifiers does. Accordingly, we have designed different domain adaptation techniques to face this problem, all of them integrated in a same framework (V-AYLA). We have explored different methods to train a domain adapted pedestrian classifiers by collecting a few pedestrian samples from the target domain (real world) and combining them with many samples of the source domain (virtual world). The extensive experiments we present show that pedestrian detectors developed within the V-AYLA framework do achieve domain adaptation. Ideally, we would like to adapt our system without any human intervention. Therefore, as a first proof of concept we also propose an unsupervised domain adaptation technique that avoids human intervention during the adaptation process. To the best of our knowledge, this Thesis work is the first demonstrating adaptation of virtual and real worlds for developing an object detector. Last but not least, we also assessed a different strategy to avoid the dataset shift that consists in collecting real-world samples and retrain with them in such a way that no bounding boxes of real-world pedestrians have to be provided. We show that the generated classifier is competitive with respect to the counterpart trained with samples collected by manually annotating pedestrian bounding boxes. The results presented on this Thesis not only end with a proposal for adapting a virtual-world pedestrian detector to the real world, but also it goes further by pointing out a new methodology that would allow the system to adapt to different situations, which we hope will provide the foundations for future research in this unexplored area.

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: Recognizing object categories in real world images is a challenging problem in computer vision. The deformable part based framework is currently the most successful approach for object detection. Generally, HOG are used for image representation within the part-based framework. For action recognition, the bag-of-word framework has shown to provide promising results. Within the bag-of-words framework, local image patches are described by SIFT descriptor. Contrary to object detection and action recognition, combining color and shape has shown to provide the best performance for object and scene recognition.
In the first part of this thesis, we analyze the problem of person detection in still images. Standard person detection approaches rely on intensity based features for image representation while ignoring the color. Channel based descriptors is one of the most commonly used approaches in object recognition. This inspires us to evaluate incorporating color information using the channel based fusion approach for the task of person detection.
In the second part of the thesis, we investigate the problem of object detection in still images. Due to high dimensionality, channel based fusion increases the computational cost. Moreover, channel based fusion has been found to obtain inferior results for object category where one of the visual varies significantly. On the other hand, late fusion is known to provide improved results for a wide range of object categories. A consequence of late fusion strategy is the need of a pure color descriptor. Therefore, we propose to use Color attributes as an explicit color representation for object detection. Color attributes are compact and computationally efficient. Consequently color attributes are combined with traditional shape features providing excellent results for object detection task.
Finally, we focus on the problem of action detection and classification in still images. We investigate the potential of color for action classification and detection in still images. We also evaluate different fusion approaches for combining color and shape information for action recognition. Additionally, an analysis is performed to validate the contribution of color for action recognition. Our results clearly demonstrate that combining color and shape information significantly improve the performance of both action classification and detection in still images.

Abstract: In this dissertation, we explore different ways to perform human action recognition in video clips. We focus on data efficiency, proposing new approaches that alleviate the need for laborious and time-consuming manual data annotation. In the first part of this dissertation, we start by analyzing previous state-of-the-art models, comparing their differences and similarities in order to pinpoint where their real strengths come from. Leveraging this information, we then proceed to boost the classification accuracy of shallow models to levels that rival deep neural networks. We introduce hybrid video classification architectures based on carefully designed unsupervised representations of handcrafted spatiotemporal features classified by supervised deep networks. We show in our experiments that our hybrid model combine the best of both worlds: it is data efficient (trained on 150 to 10,000 short clips) and yet improved significantly on the state of the art, including deep models trained on millions of manually labeled images and videos. In the second part of this research, we investigate the generation of synthetic training data for action recognition, as it has recently shown promising results for a variety of other computer vision tasks. We propose an interpretable parametric generative model of human action videos that relies on procedural generation and other computer graphics techniques of modern game engines. We generate a diverse, realistic, and physically plausible dataset of human action videos, called PHAV for “Procedural Human Action Videos”. It contains a total of 39,982 videos, with more than 1,000 examples for each action of 35 categories. Our approach is not limited to existing motion capture sequences, and we procedurally define 14 synthetic actions. We then introduce deep multi-task representation learning architectures to mix synthetic and real videos, even if the action categories differ. Our experiments on the UCF-101 and HMDB-51 benchmarks suggest that combining our large set of synthetic videos with small real-world datasets can boost recognition performance, outperforming fine-tuning state-of-the-art unsupervised generative models of videos.