mit-han-lab/spvnas
A neural architecture search method for discovering efficient 3D sparse convolution architectures used in point cloud semantic and panoptic segmentation.
Not currently ranked — collecting fresh signals.
star history
SPVNAS introduces sparse point-voxel convolution and applies neural architecture search to find efficient 3D deep learning architectures optimized for point cloud processing. The method achieved state-of-the-art results on SemanticKITTI and won challenges on NuScenes for LiDAR semantic and panoptic segmentation. Built with PyTorch, torchsparse, and torchpack for efficient sparse tensor operations.
Frequently asked
- What is mit-han-lab/spvnas?
- A neural architecture search method for discovering efficient 3D sparse convolution architectures used in point cloud semantic and panoptic segmentation.
- Is spvnas open source?
- Yes — mit-han-lab/spvnas is open source, released under the MIT license.
- What language is spvnas written in?
- mit-han-lab/spvnas is primarily written in Python.
- How popular is spvnas?
- mit-han-lab/spvnas has 622 stars on GitHub.
- Where can I find spvnas?
- mit-han-lab/spvnas is on GitHub at https://github.com/mit-han-lab/spvnas.