Meetup #7 - virtuální validace pro autonomní řízení Meetup Meetup
Public group
Location image of event venue


Sedmý ze série Meetupů komunity kolem veletrhu práce Místo konání bude ve firmě Valeo, v jejich vývojovém centru Valeo v Praze-Strašnicích (Sazečská 247/2,[masked] Praha)


1) Ground Truth Extraction from Real World Data - Pavel Jiroutek (Software Team Leader, Valeo R&D centrum)
We are currently witnessing tremendous advances in computer vision systems, that are used to teach self-driving cars how to orient themselves in an environment and make the right driving decisions. Such systems rely on learning from and validation on huge amounts of annotated visual data: that is, images from different traffic situations, even those that are only rarely encountered in real-life. My talk will focus on the potential for the use of computer generated photorealistic images to get hold of such data. Photorealistic computer graphics has become a natural part of our lives: we get to see it in essentially every Hollywood movie, but also in advertisement, architecture visualization, etc. I will discuss how generating such photorealistic images is related to accurate physically-based simulation of the behavior of light, a computational problems with its root in the development of first nuclear weapons in the Manhattan project. I will make a loop back to self-driving cars by discussing the various situations in which photorealism is crucial for proper training of the computer vision systems.

2) From nuclear reactors to pretty pictures and self-driving cars - doc. Ing. Jaroslav Křivánek, Ph.D (Docent Matematicko-fyzikální fakulty Univerzity Karlovy / Render Legion)
The process of ground truth labeling from a set of reference measurements is inevitable, yet demanding and expensive task within the development of highly automated and autonomous driving (HAD/AD) systems. For HAD, the validation cost is comparable to the cost of the development itself, and for fully AD, the costs can be even higher. Limited or inappropriate tools vastly reduce efficiency of the labeling process. Deploying automation in the labeling process can therefore save a significant amount of resources, which would otherwise be spent on extensive manual repetitive activities. The presentation will illustrate an effective annotation process that enables automation on different levels. The presented approach is based on profound know-how of the challenges the annotation process brings when used for statistical validation of HAD/AD systems combined with the utilization of the state of art machine learning techniques. Next to the ground truth extraction from physical sensors, it will be also discussed, how sensor simulation can be utilized as a virtual validation method, to extend available data set for ADAS system validation.