PoinTr

Why we included this project

3D scanners rarely capture a complete object, and PoinTr exists to fill in the missing pieces. The model treats an unordered point cloud as a sequence of point proxies with position embeddings and pushes them through a transformer encoder-decoder, which makes it a handy reference for applying transformers beyond text and images. Alongside the core weights, the repository ships pretrained checkpoints on several benchmarks, the improved AdaPoinTr variant, and a SnowFlakeNet implementation, so you get a ready baseline and a way to study how the approach evolved. Setup covers install scripts for CUDA extensions and instructions for the ShapeNet-55/34 datasets the authors introduced, letting researchers reproduce published numbers or adapt the architecture to their own pipeline without rebuilding the tooling.