Computer Vision News - April 2019
14 Challenge and Workshop Computer Vision News Ilke Demir , Kathryn Leonard, Géraldine Morin and Camilla Hahn are hosting a workshop at CVPR 2019 – which takes place in Long Beach, California in June – about the use of deep learning approaches for geometric shape understanding. Deep learning approaches for segmentation, recognition and detection tasks are already very popular; but general shape understanding tasks, such as how we can extract abstractions of shapes, have not been explored to the same extent. Challenge In the lead up to the workshop , the team are running a challenge looking at one general shape understanding task: skeleton extraction from a given shape . The communities of skeletal modelling and deep learning don’t have a huge amount of overlap, so this is something that has not been done before. There are existing methods for extracting skeletons from shapes; however, there are drawbacks to those approaches in the sense that they end up producing branches that capture meaningless information such as pixelation and other kinds of noise on the boundary. The team are hoping to circumvent this by using deep learning to extract clean skeletons . This will be a challenge because extracting a skeleton is not a robust process. If you have a noisy shape, the skeleton is going to be very different. Small changes on the boundary can produce large changes to the skeleton. These geometric shapes have a particular structure that is embedded in the skeleton. The team are proposing to explore the problem in three domains: image, point, and parametric representation. They just released a dataset paper with some very simple baselines for participants on how to approach the challenge, and people can go from there and develop their own solutions. If the challenge itself isn’t enough to tempt people to join, there are prizes on offer! The winner in each track will receive a Titan GPU sponsored by NVIDIA . Where did it all begin? The idea for the challenge was conceived when the team met at a workshop last summer in Germany, General shape understanding tasks have not been explored to the same extent as segmentation, recognition and detection tasks...
Made with FlippingBook