Masked Gamma-SSL: Learning Uncertainty Estimation via Masked Image Modeling
Overview:
This method extends the work in GammaSSL by encorporating foundation models (DINOv2) and masked image modeling. Foundation models possess a general representation which is highly suited to performing distributional uncertainty estimation, but which suffers when fine-tuning takes place. This method provides a way of both finetuning a foundation model to perform a specific task, while keeping the quality of uncertainty estimation high. Masked image modeling is used as a simpler, less hyperparameter-dependent alternative to the augmentation task in GammaSSL.
Link to $\mathrm{ar\chi iv}$ and $\mathrm{code}$.
BibTeX:
@article{williams2024mgssl,
author = {Williams, David and Gadd, Matthew and Newman, Paul and De Martini, Daniele},
title = {Masked Gamma-SSL: Learning Uncertainty Estimation via Masked Image Modeling},
journal = {IEEE International Conference on Robotics and Automation (ICRA)},
year = {2024},
}