Point Transformer
This work proposes SortNet, as part of the Point Transformer, which induces input permutation invariance by selecting points based on a learned score, to extract local and global features and relate both representations by introducing the local-global attention mechanism.
Abstract
In this work, we present Point Transformer, a deep neural network that\noperates directly on unordered and unstructured point sets. We design Point\nTransformer to extract local and global features and relate both\nrepresentations by introducing the local-global attention mechanism, which aims\nto capture spatial point relations and shape information. For that purpose, we\npropose SortNet, as part of the Point Transformer, which induces input\npermutation invariance by selecting points based on a learned score. The output\nof Point Transformer is a sorted and permutation invariant feature list that\ncan directly be incorporated into common computer vision applications. We\nevaluate our approach on standard classification and part segmentation\nbenchmarks to demonstrate competitive results compared to the prior work. Code\nis publicly available at: https://github.com/engelnico/point-transformer\n