June 2020
tl;dr: Visibility augmented deep voxel representation, with occupancy grid feature map.
Representing point cloud as xyz fundamentally destroys the difference of free space and uncertainty space (both contains no lidar points). Convolution cannot differentiate such difference based on such a representation.
WYSIWYG adds a occupancy grid feature map to the existing feature map, very much like coord conv and cam conv.
- Fast ray casting via voxel traversal --> visibility volume
- Visibility over multiple ldiar sweeps: bayesian filtering
- data augmentation
- Naive data augmentation: copy and paste rare objects such as buses. But it violates visibility rules.
- Visibility aware data augmentation
- The idea of ray casting is widely used in generating lidar simulated data, such as lidar sim.
- Review about data representation
- Talk at CVPR 2020
- github code. The codebase is based on SECOND, similar to pointPillars.