A new preprint on arXiv, titled "Pixels to Tokens: Token Space Efficient Active Learning for Low-Budget Semantic Segmentation," argues that current active learning (AL) strategies for vision transformers (ViTs) and foundation models operate at the patch or region level. The authors contend that this level of annotation may be inefficient when budgets are tight.
The paper proposes moving AL to token space, where annotations are selected based on individual tokens rather than larger spatial regions. This is positioned as a more efficient alternative for low-budget scenarios, though the abstract is truncated and does not yet include experimental results or comparisons.
As a preprint, the work has not been peer-reviewed. The full methodology and evidence will require reading the complete paper, but the core idea is a shift in granularity for active learning in semantic segmentation with modern transformer-based backbones.