Computer Vision News - August 2022

33 Learning Shape Reconstruction... us to work with different kinds of sparse training data inputs and reconstruction target resolutions. It essentially allows us to unify these different kinds of discrete measurements that we have. ” Although the training data is sparse, and the network has never seen a structure’s full shape, the model can reconstruct smooth and natural-looking shapes from just three slices . Does Tamaz think this is the aspect of the work that stood out for the judges? “ When you first look at it, it seems impossible, and it’s a bit of a surprise that you can train this model on such sparse data – but it works! ” Tamaz reveals. “ That is likely what caught people’s attention. I didn’t believe it myself at first. I tried many things to check if it was doing what it looked like it was doing. Luckily, all the experiments were successful. It was segmentations with large slice distances . If you are given three orthogonal slices, it is impossible to know what the output should be if you do not know beforehand what you are looking at. The measurement on its own is insufficient to reconstruct the shape. Take a lumbar vertebra, for example. A radiologist would know what lumbar vertebrae look like in general and could guess the shape. This model tries to mimic that behavior with a neural network learning the distribution of vertebrae shapes that naturally occur and using this knowledge to help perform these reconstructions. “ We leverage so-called implicit functions in this work, and the special property of this shape representation is that it is continuous and not discrete like meshes or voxel-based representations, ” Tamaz explains. “ This continuous representation helps