Face alignment is the process of determining a face shape given its location and size in an image. It’s used as a basis for other facial analysis tasks, human-machine interaction, and augmented reality applications. It’s also a challenging problem due to the extremely high variability in facial appearance affected by many external (illumination, occlusion, head pose) and internal factors (race, facial expression).
This session will present the latest regression-based approaches, recent training techniques, and a benchmark comparison of the most successful methods.
Speaker: Ivan Gogić, Director of Research and Development
Deep neural networks are the backbone of important algorithms implemented in the SDK for tracking, analyzing and recognizing faces developed at Visage Technologies. Since the SDK is designed to run efficiently on a variety of platforms (from Windows, Linux, and Mac OS to Android, iOS, and a variety of embedded systems), we need adequate inference engine mechanisms to enable deep network execution. Examples of such mechanisms are OpenVINO, TFLite, NCNN, and Visage Technologies’ in-house inference engine.
This session will provide an overview of inference mechanisms and our experience of their integration under a common API.
Speakers: Nikola Mrzljak & Jure Pajić, R&D Engineers
How can we ensure that a new iteration of a tracking algorithm is better than the previous one? Moreover, how can we automate the process to free up the resources for developing new features?
Learn how we do this and more in Visage Technologies in this session.
Speaker: Petar Stojanac, Head of Face Technology Division
As a part of the project with a pharmaceutical company, Visage Technologies created Maria, a virtual doctor. Maria is essentially a chatbot that moves, speaks and reacts to your emotions and your body mass index estimation in a predefined way. This means that she is more of an NPC in a video game than an intelligent agent. However, she is very flexible in what she does, which can be summarized as content-agnostic, question-answer based conversations. This was achieved via a configurable architecture that ranges from a set of predefined parameters such as conversation flow and content and body animation of the agent, all the way to automatically generated assets such as voice and mouth animation.
This presentation will provide an architectural overview of the system for conversation generation.
Speaker: Vjekoslav Ranogajec, Director of Solutions and Custom Development
Tensor-based algorithms are increasingly finding significant applications in machine learning, particularly computer vision, somewhat because they offer a structure-exploiting approach. Many of those are based on tensor contractions and different tensor factorizations.
In this session, we will introduce the basic definitions and operations, present several factorizations with applications such as face recognition and shadow reduction, and set up a motivation for more advanced uses in the field of neural networks.
Speaker: Lana Periša, R&D Engineer
One of the challenges we encountered in developing a vision system was working with data collected in China. While the numerical results of development can be used without limitations, that is not the case with images and video. Constraints on usage include exporting overly-long footage, exporting footage containing important landmarks, or exporting footage at all.
This presentation describes some of those constraints and, where possible, the solutions that enabled us to deal with them.
Speakers: Vedran Pamuković, Senior R&D Engineer & Filip Soldan, Software Engineer in Test
During the Student Internship of 2020, six students from the University of Zagreb banded together with mentors from Visage Technologies to tackle five computer vision challenges in the automotive industry. They tackled Kalman Filters, Depth Estimation, Generative Adversarial Networks, Data Reduction, and Data Augmentation and had a great time in the process. Each of them successfully completed their challenge and presented it to the entire Visage Technologies team at the end. Find out more about their challenges and results in this presentation.
Speaker: Anamarija Čavka, Data and Machine Learning Specialist
Shopping is increasingly moving online, and 2020 certainly gave it an additional push. To help brands keep up with new habits and expectations of their customers, we developed makeup|SDK – the technology that lets users virtually try on makeup in real time, just like a mirror. This presentation will explain why we decided to create a new SDK, how we managed to develop this new product in only 3 months, and what we hope to achieve in the future.
Speaker: Lea Alfier, Software Engineer