Conferences and supporting programme
Deep Learning Versus Rule-based-configurable Vision Software on Embedded Devices
This paper’s goal is to compare the use of “traditional” rule-based-configurable software with deep-learning-based approaches on embedded hardware. To facilitate this, both techniques will be presented with their respective advantages and drawbacks. Deep Learning, for example, is already being used in many applications, such as autonomous driving and smart city scenarios. Lengthy manual handcrafting of features is not necessary anymore, due to automatic feature extraction, saving time and costs. A major disadvantage are the high system requirements, as deep learning is resource-hungry, requiring specialized hardware (see the NVIDIA Jetson TX2 and the new Intel OpenVino with Movidus) More classical approaches, such as rule-based-configurable software, do not require a training phase beforehand, as rules are set up on a case-by-case basis. This makes the application more transparent, which is ideal, e.g., for validations in the medical sector. Efficiency is another benefit of this method, as the system requirements are quite low and there is no need for a vast image database before creating the application. The paper will then move on to a discussion of the presented aspects of both technologies and will conclude with the proposal of a decision tool to choose the right method for one’s application. To sum it up, there are some applications which might exclusively be solved with deep learning in the future, while others will still rely on traditional methods.
--- Date: 26.02.2019 Time: 2:30 PM - 3:00 PM Location: Conference Counter NCC Ost