This is a Pytorch implementation of our Sparse PointPillars work presented at the Sparse Neural Networks 2021 Workshop.

This repo is built on top of second.pytorch, the official codebase of the 3D detection model SECOND and the codebase used for the official implementation of PointPillars, which has been updated to provide a sanctioned implementation of PointPillars.

This codebases offers sparse backbones implemented using either a fork of spconv or Minkowski Engine sparse operations. The implementation of the COO format Pillar Feature Net is found in second/pytorch/models/pointpillars.py, the spconv backbones are found in second/pytorch/models/rpn_spconv_sparse.py, and the Minkowski Engine backbones are found in second/pytorch/models/rpn_mink_sparse.py.

To reproduce our development environment, we provide both Docker files for CUDA 10 and 11 (inside docker/) as well as a conda environment for CUDA 11 (environment.yml). When using a conda environment, the root of the repo must be added to your PYTHONPATH.