Competitive box-prompt derivation from μSAM decoder predictions

Develop a strategy to derive high-quality, over-sampling box prompts from the μSAM segmentation decoder predictions that achieves instance segmentation performance competitive with the APG point-prompt approach.

Background

APG currently derives point prompts from μSAM decoder predictions and achieves strong results, but incorporating box prompts could further improve performance because box prompts often outperform point prompts in SAM-style models.

The authors attempted to derive box prompts directly from μSAM decoder outputs but were unable to find a competitive strategy, indicating a concrete unresolved technical challenge.

References

However, we found that deriving a set of high-quality prompts that over-sample objects was challenging and did not yet find a strategy competitive with the simpler point prompt derivation.

— Revisiting foundation models for cell instance segmentation  (2603.17845 - Archit et al., 18 Mar 2026) in Discussion