Computer Vision News - April 2019

12 Computer Vision News Project Management Tip Management Let’s see what happens when data is collected during an on-going project; in other words, when input images or videos keep coming at the same time as you develop your network. In such a case you might find yourself in a feedback loop, requesting for specific data and changing your network as the data grows. Two main things that would help you with this would be flexibility in network architecture, and a careful follow up of the training procedures. We’ll start with code flexibility: Every deep learning development project implies experimenting with varied deep learning methods and tricks. Therefore, you’d like to be able to easily change network architecture, without the need to modify your entire previously written procedure. When data is growing, it might be that the project is still changing its shape. Some of its challenges are yet undiscovered and perhaps the network’s objective will change a few times before it will mature. In addition to changes in the internal architecture of the network, this could entail changes of network interface: the loss function, the inputs and outputs, and correspondingly the format of the ground truth. You want to be able to quickly go back and forth between network architectures in order to lunch trainings and compare their results. Thus, in order to make life easier, consider a code structure that allows flexible network interface, a generic prediction output, and a modular network architecture. The process of better understanding the network’s problems and objective could be efficiently driven by iteratively training the network on the existing data, and then inspecting its false predictions. By analyzing the resulting errors, you could identify specific cases in which the network is not performing well. What common features and characteristics do these cases share? This could teach you what the network has actually learnt, and what are the challenges it is still facing. At early steps of the project you might discover that the network hasn’t learnt anything sophisticated, but it is predicting based on degenerated features that happen to fit your ground truth. In this case, this means that your data is not varied enough with regard to the degenerated feature, and it is not reliable enough in its representation of When Data is Collected During the Project 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 Yael Zak tells us about another aspect of Project Management with Deep Learning: When Data is Collected During the Project . It’s another handy tip by RSIP Vision for Project Management in Computer Vision . “Data is everything in this game!”