Computer Vision News - January 2019

10 Computer Vision News Project Management Tip Management Last month’s lecture dealt with one of the main issues that a project manager must solve: collecting and selecting data for medical projects . This month we discuss about how this data should be annotated . This too is a critical phase in deep learning based projects . The easier case is when the client requesting the work has available annotated data for the project. Another optimal case is when open datasets are available to use. When it happens, there is no need for further data annotation. Still, this is not the common case, since labeling medical data is at least as problematic as finding it . Moreover, we have seen that training data should be as diverse as possible, to cover as many cases as possible in the real world. This adds further complexity to the annotation task. Greater data diversity generally entails the need for better trained annotators, with advanced medical expertise. When the company has its own in- house doctor, the solution of this problem is much easier: the doctor can provide these annotations or at least supervise them and improve their accuracy. When limited medical eyes are available, the company needs to think of creative ways to secure accurate labels for its medical images. One of the possible solutions is the use of conventional (not deep learning) semi-automated computer vision techniques to generate these labels. This is what Ron Soferman calls “ computers teaching computers ”. Indeed, it may happen that a previous version of the software, though slower and less precise, will be able to run offline and provide sufficiently accurate annotations . After some necessary corrections, these initial results will turn into accurate and reliable training data. When even a previous version of the software is not available, it is still possible to write a simple program that supports the annotation task: for example, a simple threshold of the image sometimes suffices to provide a partial segmentation. This method will generate much noise, but at the same time it may simplify the task of the human annotator. In same cases, this hint can save up to 80% of the labeling effort. Annotating Data for Medical Projects RSIP Vision’s CEO Ron Soferman has launched a series of lectures to provide a robust yet simple overview of how to ensure that computer vision projects respect goals, budget and deadlines. This month Arik Rond tells us about Annotating Data for Medical Projects . It’s another tip by RSIP Vision for Project Management in Computer Vision . “…computers teaching computers…”