Mitigating Distributional Shift in Semantic Segmentation via Uncertainty Estimation from Unlabelled Data

Overview:

It is important that methods used in safety-critical settings know when they are likely to fail. Therefore, uncertainty estimation in deep learning is a key field for applications such as robotics. This method, named GammaSSL, trains a segmentation network that can estimate when it is likely to fail as a result of distributional shift - i.e. performs distributional uncertainty estimation. It achieves this by using uncurated, unlabelled out-of-distribution training images.

Link to $\mathrm{ar\chi iv}$ and $\mathrm{code}$.

BibTeX:

@article{williams2024gammassl,
  author    = {Williams, David and De Martini, Daniele and Gadd, Matthew and Newman, Paul},
  title     = {Mitigating Distributional Shift in Semantic Segmentation via Uncertainty Estimation from Unlabelled Data},
  journal   = {IEEE Transactions on Robotics (T-RO)},
  year      = {2024},
}

Paper: