The rapid development of virtual reality (VR) and 3D sensors call for better algorithms for object localization methods to enhance the user’s 3D experience. Size estimation is the problem of predicting the dimensions of an object in a 3D scan, which is necessary for the VR experience as it is difficult to judge the real-world size in a VR environment. Accurate size estimation for the walls and furniture can also expose additional information on virtual tours of houses or museums. Augmented reality (AR) is used widely in indoor tourism and virtual tours. Today, state-of-the-art 3D semantic and instance segmentation models have demonstrated impressive performance for their respective tasks. There are two clustering-based approaches to produce more accurate size estimations from these models’ predictions in this paper. One of them can segment and measure the sizes of walls given the points labeled as walls. The other algorithm refines the instance proposals from a segmentation model by removing mislabeled noise. Compared to a baseline method of finding the tightest bounding box of SoftGroup’s [1] instance predictions on the S3DIS [2] dataset, with a 2% error threshold, our implementations can achieve +8%, +6%, +4% higher accuracies on wall length, chair length, and chair width measurements respectively.