Computer Vision News - February 2019

16 Computer Vision News Challenge algorithm that is independent and different for every single task. It would be ideal if there was an algorithm that performs very well on many different tasks at the same time. This was really the primary aim of the challenge. Can an algorithm, without any human interaction, learn many different things even when the tasks are extremely different from each other? The tasks could be anything from segmenting tumors from CT, segmenting the hippocampus from the MR... pretty much every single semantic segmentation task you could possibly imagine with very large and very small objects, with a large number of datasets and very small numbers of datasets. The ten tasks were purposely selected to encompass very different types of problems that you see in the medical imaging domain. What we found in the end is that yes, indeed, the better algorithms, the ones that won the challenge, were able to perform very well in a large amount of problems. They were actually able to achieve performance very close to state-of-the-art on a large number of problems. The methods that were proposed were general purpose. They were not task- specific. They were not designed specifically to solve a task. They were designed to solve many possible tasks with very general purpose. They should perform pretty much as well as algorithms tailored and designed for a specific task. Even more interesting than that, the winning algorithm, the algorithm from DKFZ, got state-of-the-art performance in two tasks. Even though it was not developed on purpose for those two tasks, they performed better than the best algorithm until that day on that specific task. They had state-of-the-art performance just behind the submissions of that specific year from NVIDIA which was quite phenomenal. They could demonstrate that you can have algorithms that are general, but also perform extremely well. It also demonstrates that you can get state- of-the-art performance on certain tasks even though those algorithms are general purpose. Something that needs to be discussed and we need to understand is when we talk about generalizing these algorithms, we talk about algorithms as general learning systems, but they Challenge