Alternative representation anchors and geometric targets for GAFT

Investigate alternative representation anchors and more accurate geometric targets for Geo-Anchored Fine-Tuning (GAFT) as unresolved design choices in geometry-guided adaptation for RGB hazard identification.

Background

The paper evaluates GAFT with a KL-divergence representation anchor that discourages the adapted DINOv2 backbone from deviating substantially from the frozen pretrained representation. Its geometric supervision uses a proximity-weighted geometric proxy derived from RGB–depth traversal data and a fitted ground plane, rather than a true hazard mask. The authors explicitly identify alternative anchoring mechanisms and more accurate geometric targets as remaining open design choices, leaving their comparative effectiveness unresolved.