ICCV Daily 2021 - Wednesday

“ We improve those metrics for semantic segmentation, especially in the urban sense. Our work starts from the intuition that the distribution of prediction scores is significantly different from each other for predicted classes. By normalizing them to have the same semantics, that’s the starting point of our work. ” There are three main contributions of this work. The first is trying to align the differently formed prediction scores , which is very commonly observed in semantic segmentation. Previous work didn’t tackle this problem. “ The second contribution is that in semantic segmentation, the model detects the in-distribution pixels as the out-of-distribution pixels in the boundary regions, ” Jungsoo explains. “ One of the reasons is because the boundary regions are where the classes change, like from cars to road. Because of that characteristic, there are a lot of false positives occurring in the boundary regions. We try to reduce the false positives with a module named boundary-aware pooling. For the third contribution, there are some noises still existing in semantic segmentation after applying the first and second parts, so we reduce those noises using a Gaussian filter. ” 19 DAILY ICCV Wednesday Sanghun Jung and Jungsoo Lee