Unsupervised Multi-Task Feature Learning on Point Clouds

3D point clouds are used in AEC applications such as Scan2BIM and also are compact representations of complex 3D models such as models used in manufacturing. Annotating point clouds is a labor-intensive and time-consuming task that is required to accurately classify or segment them. To address this, we introduce a unsupervised multi-task approach to learn high level features which in turn minimizes the need for labels. This can help to automatically segments objects to their parts or cluster 3D models. Specifically, automatic segmentation of 3D models to their constituent parts can help to automatically expand “Design Graph.”

Abstract

Unsupervised Multi-Task Feature Learning on Point Clouds

Kaveh Hassani, Mike Haley

IEEE International Conference on Computer Vision (ICCV) 2019

We introduce an unsupervised multi-task model to jointly learn point and shape features on point clouds. We define three unsupervised tasks including clustering, reconstruction, and self-supervised classification to train a multi-scale graph-based encoder. We evaluate our model on shape classification and segmentation benchmarks. The results suggest that it outperforms prior state-of-the-art unsupervised models: In the ModelNet40 classification task, it achieves an accuracy of 89.1% and in ShapeNet segmentation task, it achieves an mIoU of 68.2 and accuracy of 88.6%.