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

Official checkpoints for the ECCV 2026 MedVidU Workshop paper RefineRank.

RefineRank couples two frozen backbones — a MedVLM (Qwen2.5-VL architecture) and GroundingDINO — with a compact 1.25M-parameter trainable module, RefineNet (QueryConditionedProposalAdapter). RefineNet uses MedVLM language and regional features to predict coordinate corrections and box-quality scores for GroundingDINO proposals; a parameter-free decoder then returns the highest-scoring original or refined box.

• Code, training & inference guide: https://github.com/linzhe001/RefineRank
• Headline result: 0.421 STG mIoU on the archived MedVidBench Community leaderboard snapshot (27 July 2026) — the best STG mIoU among the ten ranking metrics on that snapshot.
• Controlled evaluation (video-separated split over CholecTrack20 / CoPESD / EgoSurgery): STG mIoU 0.2719 → 0.4534 over the frozen MedVLM + GroundingDINO baseline.

Contents

This repo hosts the complete checkpoints/ tree expected by the code repository — three flat folders, each holding its core files directly:

checkpoints/
├── vlm/                                  # frozen MedVLM, HF format (~16 GB)
│   ├── config.json, generation_config.json, tokenizer*, preprocessor_config.json, ...
│   └── model-00001..00004-of-00004.safetensors
├── grounding_dino/
│   └── groundingdino_swinb_cogcoor.pth   # frozen GroundingDINO SwinB (~895 MB)
└── refinenet/                            # trained RefineNet, this work (~5 MB)
    ├── proposal_adapter_full.pt          # SHA-256: 932e479b…463d9b
    └── deployment_manifest.json

• refinenet/proposal_adapter_full.pt is the exact checkpoint behind the paper's MedVidBench submission (run_iter132_submission). SHA-256: 932e479b854c3d5fbafee25a3fcf9e6481e864fe98a6867502d6ebff39463d9b.
• vlm/ and grounding_dino/ are third-party frozen weights (uAI-NEXUS-MedVLM by UII-AI and GroundingDINO by IDEA-Research), mirrored here for one-stop reproducibility. They are never fine-tuned by RefineRank; please follow their original licenses and cite the original works.

Usage

pip install "huggingface_hub[hf_transfer]"   # hf_transfer optional, faster
hf download linzher/RefineRank --local-dir .  # restores the checkpoints/ tree

git clone https://github.com/linzhe001/RefineRank
cd RefineRank
pip install -r requirements.txt
# place the downloaded checkpoints/ next to interface.py, then:
python interface.py predict   # auto-discovers checkpoints/refinenet/

Citation

@inproceedings{jiang2026refinerank,
  title     = {RefineRank: Joint Box Refinement and Ranking for Surgical
               Spatio-Temporal Grounding},
  author    = {Jiang, Linzhe and Huang, Jiayuan and Zhang, Changhao and
               Jiang, Chunyang and Mao, Zhehua and
               Garcia-Peraza-Herrera, Luis C. and Hoque, Mobarak I.},
  booktitle = {ECCV Workshops (MedVidU)},
  year      = {2026}
}