According to an aspect, a method for creating a three-dimensional map for augmented reality (AR) localization includes obtaining a digital representation of a scene of an AR environment, where the digital representation has been captured by a computing device. The method includes identifying, using a machine learning (ML) model, a region of the digital representation having visual data identified as likely to change, and removing a portion of the digital representation that corresponds to the region of the digital representation to obtain a reduced digital representation, where the reduced digital representation is used to generate a three-dimensional (3D) map for the AR environment.