(English version at the bottom)
Tive o prazer de bater um papo com Renan Wille, engenheiro eletricista e especialista em visão computacional. Durante nossa conversa, discutimos sua trajetória profissional e alguns dos projetos nos quais ele trabalhou, além dos desafios em desenvolver soluções de visão computacional.
I had the pleasure of chatting with Renan Wille, an electrical engineer and computer vision specialist. During our conversation, we discussed his professional career and some of the projects he has worked on, as well as the challenges of developing computer vision solutions.
At the beginning of his university studies, Renan soon realized that there were two main branches of electrical engineering: power energy and telecommunications. However, neither of these really caught his interest. Due to family influences, Renan grew up immersed in the world of professional photography, developing a deep interest in the technique and physics behind this art. In addition, technology, and more specifically computing, also piqued his curiosity. By uniting these two fields, photography and computing, Renan found his true passion: computer vision. He had the opportunity to explore this field for the first time during a scientific initiation project in 2010.
The project, supervised by Dr. Giselle L. Ferrari, lasted two years and aimed to develop a pupilometer for biomedical use. This system included both the physical part, i.e. the hardware and electronics, and the embedded software, and its purpose was to measure how the pupil of the eye dilated in response to the brightness of a flash of light. To do this, an LED flash light was used to stimulate the eye, while a video camera captured the reaction of the pupil.
The video recording started two seconds before the flash, which lasted ten milliseconds, and continued until three seconds after the flash. With the help of a custom algorithm, the system automatically analyzed the data collected. It calculated the response time and the delay until the pupil began to contract, the delay until the pupil reached its maximum constriction and the amplitude of the pupillary reflex. This project has been presented at technical seminars and has received recognition, including a first place award at scientific initiation events. One of the biggest challenges encountered during the development of the project was developing the complete system, from the electronic design to the embedded software that processes the images to obtain the pupil response curves.
Halfway through college, Renan had another great opportunity: he was accepted to take part in a sandwich graduation program, which meant spending around a year and a half studying at a technical institution in France with an internship and scholarship. Besides gaining a deeper understanding of the new culture and enhancing his proficiency in the new language, this experience also led to substantial progress in his technical education. The more practical approaches to teaching were particularly valuable. Renan had the chance to attend classes in laboratories equipped with state-of-the-art technology. This immersion allowed him to gain a deeper understanding of some important electronic engineering concepts, such as the transmission of electromagnetic waves.
After returning to Brazil, Renan finished his degree and soon entered the job market. After a first internship at another company, Renan joined the company where he is still working today. This company develops, among other things, special devices for capturing, processing and identifying license plate images using optical character recognition. This is also an area of application in the field of computer vision.
It's important to note that Renan graduated in 2014, i.e. during the early years of the revolution and spread of machine learning models. In 2012, for example, the AlexNet scientific paper marked a significant advance in the image recognition capabilities of convolutional neural networks. In addition, two of the most popular libraries specializing in solving computer vision problems, TensorFlow and PyTorch, were released in 2015 and 2016, respectively. Before these innovations, it was common to use more traditional approaches to tackle such problems.
A practical example of the challenges Renan faced before this revolution was the extensive use of the vectorization technique in the mathematical operations of his models. This aimed to speed up processing and allow the CPU to perform calculations in parallel, reducing the number of instructions required. This approach was especially relevant for operations involving matrix multiplications, where vectorization could significantly improve processing efficiency on this type of hardware.
Renan is currently focused on creating robust software solutions, mainly for computer vision applications involving license plate identification. To achieve this goal, he values the application of modern software development approaches, such as the implementation of a Continuous Integration (CI) pipeline. This approach provides confidence that any changes made to the source code will meet all the planned functions and will not cause unwanted issues in the systems in production. Some of the tips he recommends for creating a robust CI process are: bring the entire software generation process into the CI, from construction to testing; include several tests for problems encountered on a day-to-day basis that have been corrected, to prevent them from occurring again; use pull requests and test the code before it enters the final repository; and add automatic tools to test for code standardization.
The field of computer vision applications is constantly expanding, with significant growth and incredible opportunities in areas such as robotics and autonomous vehicles. With his rich experiences and theoretical and practical knowledge in this domain, Renan is undoubtedly well-placed to make a significant impact and contribute to ongoing advances in this evolving sector.
Links:
> Projeto de iniciação cientifica:
Estudo de iluminação e estimulação de pupilômetro dinâmico 1
> Projeto do mestrado:
Reconhecimento de marca e modelo de veículos a partir de imagens
> PyTorch
> AlexNet
No comments:
Post a Comment