Abstract: While depth tends to improve network performances, it also makes gradient-based training more difficult since deeper networks tend to be more non-linear. The recently proposed knowledge distillation approach is aimed at obtaining small and fast-to-execute models, and it has shown that a student network could imitate the soft output of a larger teacher network or ensemble of networks. In this paper, we extend this idea to allow the training of a student that is deeper and thinner than the teacher, using not only the outputs but also the intermediate representations learned by the teacher as hints to improve the training process and final performance of the student. Because the student intermediate hidden layer will generally be smaller than the teacher's intermediate hidden layer, additional parameters are introduced to map the student hidden layer to the prediction of the teacher hidden layer. This allows one to train deeper students that can generalize better or run faster, a trade-off that is controlled by the chosen student capacity. For example, on CIFAR-10, a deep student network with almost 10.4 times less parameters outperforms a larger, state-of-the-art teacher network.

Abstract: The bag of visual words is a well established representation in diverse computer vision problems. Taking inspiration from the fields of text mining and retrieval, this representation has proved to be very effective in a large number of domains.
In most cases, a standard term-frequency weighting scheme is considered for representing images and videos in computer vision. This is somewhat surprising, as there are many alternative ways of generating bag of words representations within the text processing community. This paper explores the use of alternative weighting schemes for landmark tasks in computer vision: image
categorization and gesture recognition. We study the suitability of using well-known supervised and unsupervised weighting schemes for such tasks. More importantly, we devise a genetic program that learns new ways of representing images and videos under the bag of visual words representation. The proposed method learns to combine term-weighting primitives trying to maximize the classification performance. Experimental results are reported in standard image and video data sets showing the effectiveness of the proposed evolutionary algorithm.

Abstract: The goal of our experiment is to develop a useful and accessible tool that can be used to evaluate a patient's health by analyzing handwritten strokes. We use a cloud computing approach to analyze stroke data sampled on a commercial tablet working on the Android platform and a distant server to perform complex calculations using the Delta and Sigma lognormal algorithms. A Google Drive account is used to store the data and to ease the development of the project. The communication between the tablet, the cloud and the server is encrypted to ensure biomedical information confidentiality. Highly parameterized biomedical tests are implemented on the tablet as well as a free drawing test to evaluate the validity of the data acquired by the first test compared to the second one. A blurred shape model descriptor pattern recognition algorithm is used to classify the data obtained by the free drawing test. The functions presented in this paper are still currently under development and other improvements are needed before launching the application in the public domain.

Abstract: Finding a way to estimate quickly and robustly the pose of an image is essential in augmented reality. Here we will discuss the approach we chose in order to get closer to real time by using SIFT points [4]. We propose a method based on filtering both SIFT points and images on which to focus on. Hence we will focus on relevant data.

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: Human factors cause most driving accidents; this is why nowadays is common to hear about autonomous driving as an alternative. Autonomous driving will not only increase safety, but also will develop a system of cooperative self-driving cars that will reduce pollution and congestion. Furthermore, it will provide more freedom to handicapped people, elderly or kids.
Autonomous Driving requires perceiving and understanding the vehicle environment (e.g., road, traffic signs, pedestrians, vehicles) using sensors (e.g., cameras, lidars, sonars, and radars), selflocalization (requiring GPS, inertial sensors and visual localization in precise maps), controlling the vehicle and planning the routes. These algorithms require high computation capability, and thanks to NVIDIA GPU acceleration this starts to become feasible.
NVIDIA® is developing a new platform for boosting the Autonomous Driving capabilities that is able of managing the vehicle via CAN-Bus: the Drive™ PX. It has 8 ARM cores with dual accelerated Tegra® X1 chips. It has 12 synchronized camera inputs for 360º vehicle perception, 4G and Wi-Fi capabilities allowing vehicle communications and GPS and inertial sensors inputs for self-localization.
Our research group has been selected for testing Drive™ PX. Accordingly, we are developing a Drive™ PX based autonomous car. Currently, we are porting our previous CPU based algorithms (e.g., Lane Departure Warning, Collision Warning, Automatic Cruise Control, Pedestrian Protection, or Semantic Segmentation) for running in the GPU.