Cyril Stachniss
link : https://www.youtube.com/watch?v=vtnh9TExdOI&t=2833s
- Gaussian distribution doesn't take outliers into account
- Huber loss is better, downweights large errors
- Adaptive loss function is even better, but depend on a alpha parameter
- Optimisation pipeline that alternates iterations between alpha and whole state
- impressive results on ICP with moving objects
- Semantic maps on LiDAR data with the range image in a CNN
- SLAM with semantic ICP exploits semantic to remove or ignore moving objects