Computer Vision News - November 2016

Computer Vision News Computer Vision News Research 9 Research That was slower and worse than our final solution. We also tried with typical globalization with normalized cuts that has been there forever. Though it was very good, we finally got a better result by getting rid of globalization and doing it in a single pass. CVN: How did you go through this process? Kevis: We still didn’t have a story one month before the deadline. What was done here was work of the very last minute, but we improved quite a lot. After submission, when we look back at what we had one month before, it was nothing. All of this work was done within a very short timeframe. CVN: How were you able to succeed with such a short deadline? Kevis: Coffee! [ laughs ] CVN: Coffee is the solution for everything… [more laughs] And besides that? Kevis: Except for our interest in the work, the fact that the deadline was approaching was a good motivation for us. CVN: When you started this project, why did you decide to focus on this idea? Were you confident that you would find something in the end? Kevis: Pablo Arbeláez is the author of the original Multiscale Combinatorial Grouping. All of that MCG was really useful. The main motivation was to bring the power of deep learning into it. The second motivation was that we know that CNN has the multiscale information we need. We knew that we wanted to take out this information in a single pass of the CNN. These two driving forces bring what we know would work in MCG in deep learning, one thing is to do it in a single forward pass. These were the two main driving forces that guided us to the final solution. CVN: What makes your paper special? Jordi: In the same way that MCG was very practical for people, I think that this is going to be very useful tool for anyone in very different sets of problems. Segmentation is a very basic problem. You can base a lot of things on it. Since it’s a fast algorithm, it’s elegant in the sense that it just works in a single pass of the CNN. This simple idea will be very useful for a lot of people. I think that’s what makes it outstanding. People are going to use it so we have made the code and the pre-computed results for everything available . You can have all of the evaluation online so if anyone thinks that they want to compare to us, they can send us the results, and we will put it online. Also, for any of the images in COCO, you can look for them and see the results, ground truth, and all of the competing algorithms. If someone is working on this and wants to have the results online, they can talk to us to have it put online. Apart from the code, we have all of these pre-computed proposals, segmentations, and boxes for large data sets of whole data sets. That’s very handy for people who are on COCO. They don’t have to deal with any code or anything. We also want to convey that if anyone is working an another data set and would like to have it here, they can tell us, and we will make it available. We want this to be very useful for people and very easy to use. CVN: What is your next step? Jordi: This is just generic segmentation, and our next step would be to go into semantics: go directly into the final problem. Also, extending it to video is always good.