The RECODists Hilario Seibel Jr. and Daniele Rodrigues have earned their Ph.D. degrees last Friday September 29th at IC/Unicamp. The details are shown below:
Ph.D. candidate: Hilario Seibel
Title: Super-resolution in low-quality videos for forensics, surveillance, and mobile applications
Supervisor: Siome Goldenstein
Co-supervisor: Anderson Rocha
Super-resolution (SR) algorithms are methods for achieving high-resolution (HR) enlargements of pixel-based images. In multi-frame super resolution, a set of low-resolution (LR) images of a scene are combined to construct an image with higher resolution. Super resolution is an inexpensive solution to overcome the limitations of image acquisition hardware systems, and can be useful in several cases in which the device cannot be upgraded or replaced, but multiple frames of the same scene can be obtained. In this work, we explore SR possibilities for natural images, in scenarios wherein we have multiple frames of a same scene. We design and develop variations of an algorithm which rely on exploring geometric properties in order to combine pixels from LR observations into an HR grid; two variations of a method that combines inpainting techniques to multi-frame super resolution; and three variations of an algorithm that uses adaptive filtering and Tikhonov regularization to solve a least-square problem.
Multi-frame super resolution is possible when there is motion and non-redundant information from LR observations. However, combining a large number of frames into a higher resolution image may not be computationally feasible by complex super-resolution techniques. The first application of the proposed methods is in consumer-grade photography with a setup in which several low-resolution images gathered by recent mobile devices can be combined to create a much higher resolution image. Such always-on low-power environment requires e active high-performance algorithms, that run fastly and with a low-memory footprint.
The second application is in Digital Forensic, with a setup in which low-quality surveil- lance cameras throughout the cities could provide important cues to identify a suspect vehicle, for example, in a crime scene. However, license-plate recognition is especially di cult under poor image resolutions. Hence, we design and develop a novel, free and open-source framework underpinned by SR and Automatic License-Plate Recognition (ALPR) techniques to identify license-plate characters in low-quality real-world traffic videos, captured by cameras not designed for the ALPR task, aiding forensic analysts in understanding an event of interest. The framework handles the necessary conditions to identify a target license plate, using a novel methodology to locate, track, align, su- per resolve, and recognize its alphanumerics. The user receives as outputs the rectitude and super-resolved license-plate, richer in details, and also the sequence of license-plates characters that have been automatically recognized in the super-resolved image.
We present quantitative and qualitative validations of the proposed algorithms and its applications. Our experiments show, for example, that SR can increase the number of correctly recognized characters posing the framework as an important step toward providing forensic experts and practitioners with a solution for the license-plate recognition problem under difficult acquisition conditions. Finally, we also suggest a minimum number of images to use as input in each application.
Ph.D. candidate: Daniele Rodrigues
Title: Complex Network Measurements in Graph-based Spatio-Temporal Soccer Match Analysis
Supervisor: Ricardo Torres
Soccer match analysis is of paramount importance in the definition of appropriate training programs and game strategies. The increasing availability of sport-related data in the recent years, due to the use of modern tracking systems, has allowed advances in sports analytics, providing coaches with valuable information for match and teams analysis. The availability of these data, on the other hand, challenges science to develop tools capable of storing, visualizing, and analyzing this large volume of information. Soccer analyses are usually performed using matches’ statistics, events (e.g., passes and shots on goal) and players location data. Related studies have been representing the matches’ events as a single graph, where players are vertices and edges are actions performed among them during the match. The graph is then analyzed from a complex network measurement perspective. Although this approach provides interesting insights about the tactical actions occurred during the game, revealing some tactical patterns, it disregards the spatio-temporal aspects inherent to the sport, as the positioning of the players on the pitch, and the moment in time when relevant actions occur.
This thesis addresses these shortcomings by presenting a soccer game analysis framework. We propose a new approach for soccer match analysis, based on graphs, that considers the spatio-temporal characteristics, intrinsic to the dynamic of soccer. We propose to represent the match as a temporal graph, by encoding players’ location on the pitch into instant graphs, in which vertices represent players in their real location and edges are defined based on their distance in the field and the possibility of short pass exchanges. We demonstrate that this representation, named opponent-aware graph, which takes into account the presence of opponents, and the diversity entropy measurement are effective tools for determining the role of attacking players in a match and the probability of successful passes. By taking into account different measurements of complex networks in temporal graphs, this study also investigates the feasibility of using complex network measurements and machine learning algorithms to characterize the role of players in a match. The results allow to further characterize the decision-making process of players, providing interesting insights to coaches and researchers for possibly improving training strategies. This study also addresses the visualization of temporal graphs problem by introducing the Graph Visual Rhythm, a novel image-based representation to visualize changing patterns typically found in temporal graphs. This representation is based on the concept of visual rhythms, motivated by its capacity of providing a lot of contextual information about graph dynamics in a compact way. We validate the use of graph visual rhythms through the creation of a visual analytics tool to support the decision-making process based on complex-network-oriented soccer match analysis.