RefineRank: Joint Box Refinement and Ranking for Surgical Spatio-Temporal Grounding

This demo runs the full RefineRank pipeline: GroundingDINO generates candidate boxes, a frozen MedVLM (Qwen2.5-VL) extracts visual features, and a lightweight RefineNet (~1.25M params) refines and ranks the boxes to produce the final spatio-temporal grounding.