Computer Vision News - March 2018

This month we learn about Project Management in Deep Learning. The lecture is given by Yehiel Shilo , RSIP Vision. Yehiel earned a PhD from the Hebrew University of Jerusalem. It’s another tip by RSIP Vision about Project Management in Computer Vision . The first step needed in a computer vision project involving Deep Learning is to receive precise definitions by the client , for who the work is done: the problem at hand must be specified and a large dataset provided. Let’s discuss now the data and next month we shall discuss the code. Data received must be analyzed and understood. The project manager must assess if any pre-processing is needed for the data: that might be cropping irrelevant parts of images or videos; focusing only on relevant chunks of data might significantly reduce further workload. Depending on the type of network used, it might be needed to standardize the size of the images accordingly. Some image processing techniques might be used at this point to improve the readability of the data at hand. Some of the networks do that already, but it is advisable to help them. At this point, we add annotations to the data, sometimes via a human operator and other times via an automated procedure. It is advisable that annotation be done by different agents, so that we can assess the variability inter- and intra-operator, for instance by showing the same picture to several agents and even to the same agent several times. Data must now be split in three separate categories. The largest group is data selected to train the network . Then we have validation data , the role of which is to check during the training of the network what results will it give on an unknown portion of data. The third group is data on which the network will be tested after its optimization. The difference between validation and test data is that the former is used to check the effects of modifications to the model and to its architecture during its training. Changes will be adopted or rejected depending on the result of the validation process. The latter will be the last one to be used, in order to test the network on data which is still unknown to it. When dataset is not large enough, training data might be augmented as you can see in the video below . The project manager must now pick those methods to enrich the data which are the most realistic, so that also augmented data can be used to train the network. Project Management in Deep Learning Management Project Management Tip 27 Computer Vision News