| Description | This submission is based on our paper "Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution", which was published at ECCV 2020. We use a manually designed efficient point cloud segmentation network which is composed of multiple Sparse Point-Voxel Convolutions, as described in our paper. The solution is open-sourced at https://github.com/mit-han-lab/spvnas. In this challenge, we use a variant of SPVCNN that consumes more computation than last year and apply a longer training schedule. We also incorporate the Lovasz-Softmax loss with the standard cross-entropy loss. |