Broaden the LFPR resolution-routing policy beyond small targets

Investigate whether extending the label-free precision refinement (LFPR) resolution router beyond predicted-small bounding boxes improves grounding accuracy, given that the current restriction to the small bucket is described as a conservative operating choice rather than an empirically necessary limitation.

Background

LFPR routes only predictions whose predicted box has a geometric size below 128 pixels to a higher-resolution full-image inference. Predictions in the medium and large buckets retain the default-resolution answer. The paper initially motivated this policy by lower baseline accuracy for smaller annotated targets.

After correcting an implementation defect that had silently discarded the requested pixel budget, the authors report that resolution routing also changes performance outside the smallest target bucket: the medium bucket gains substantially at [email protected] and the large bucket gains modestly. They therefore characterize restricting the router to the small bucket as conservative rather than empirically forced, leaving open whether a broader routing policy would yield better results.