Computer Vision News - September 2018

Team - How to Split the Work in AI projects Management 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 we learn about Splitting the Work in AI Projects . It’s another tip by RSIP Vision for Project Management in Computer Vision . Split the work among the team in such a way that everyone gets an interesting role and gives a significant contribution to the project Some AI projects are big enough to require a team of 4 to 5 programmers. If the project manager wants to start the coding as soon as possible, it is necessary to allocate resources efficiently. At the same time, everybody is eager to work at the Deep Learning part, get a feel of the results and solve the problem. Here are options to split the work among the team in such a way that everyone gets an interesting role and gives a significant contribution to the project: 1) Different approaches : we can try to use different architectures for the network and even compare the work in Patch-based vs full convolutional neural network or other options. 2) Augmentation : datasets are rarely found in sufficient size and overfitting may occur even with large datasets. It is often necessary to perform data augmentation in order to obtain the needed dataset for the learning phase. 3) Annotation : sometimes it’s easier to give the work to annotators; but it is often possible to perform a semi- automated annotation using computer vision technique. The outcome of this might be increased speed and/or a larger dataset. 4) Feature extraction : this new idea is based on the thought that good features can provide a better analysis of the results. For instance, going deeper through an average result of 90%, we might find that most failure come from specific cases and features, shedding light upon the weaker points of the network or the dataset. 5) Sanity check : sometimes, when rare cases of bad example cannot be supplied to the database, we can use tools to detect failures that we do not want to transform into false positives. 6) Different Pre-processing : in many cases, pre-processing might give much better images, that can yield better results. It will be worthwhile to allocate resources in this direction as well. 7) Open source options : scientific literature provides alternatives which are offered with open source code. This enables testing different approaches in an efficient way, with the goal of finding the most valuable option. I am confident that you can successfully apply these ideas and principles! 22 Project Management Tip Computer Vision News