Problem with earlier methods - PointNet → ignores local geometry- PointNet++ → captures local structure but max pooling → too simple, loses informationThe paper proposesInstead of : neighbor features → max pooling → outputThey do : neighbor features + geometry → attention → weighted sumThe structure of RandLA-Net Local Spatial Encoding (LocSE) ① Input : (N,3+d)- N = number of points- 3 = XYZ coo..