June 2020
tl;dr: Drive once, build, and traverse multiple times to generate new data.
Use GAN to bridge the gap of surfel baseline, which is usually with gaps and edges.
Surfel is "surface element," analogous to a "voxel" (volume element) or a "pixel" (picture element).
- surfel GAN generates a photorealistic model
- lidar sim focuses on lidar data simulation, which is somewhat easier.
It can allow closed-loop evaluation of the whole AD stack.
- Build environment from a single run through a scene of interests.
- Simulate other transversals through the scene for virtual replay.
- Use a GAN model to close the domain gap between synthetic data and real data
- Uses lidar data
- Questions and notes on how to improve/revise the current work