OneRefPose – Installation & Inference Guide

---

1. 📦 Data & Pretrained Weights

1.1 Download Weights

Download pretrained models: https://drive.google.com/drive/folders/1DFezOAD0oD1BblsXVxqDsl8fj0qzB82i

Required checkpoints:

• Refiner: 2023-10-28-18-33-37
• Scorer: 2024-01-11-20-02-45

---

1.2 Directory Structure

weights/ ├── 2023-10-28-18-33-37/ # refiner └── 2024-01-11-20-02-45/ # scorer

---

Create directories:

mkdir -p weights/2023-10-28-18-33-37 mkdir -p weights/2024-01-11-20-02-45

---

1.3 Demo Data

mkdir -p demo_data/

extract demo data into demo_data/

---

1.4 LINEMOD Dataset

pip install -U "huggingface_hub[cli]"

export DATASET_NAME=lm

huggingface-cli download bop-benchmark/$DATASET_NAME
--local-dir ./${DATASET_NAME}/
--repo-type=dataset

---

2. 🛠 Installation (Conda)

2.1 Environment

conda create -n onerefpose python=3.9 -y conda activate onerefpose

---

2.2 Eigen

conda install -c conda-forge eigen=3.4.0 -y export CMAKE_PREFIX_PATH="$CMAKE_PREFIX_PATH:$CONDA_PREFIX"

---

2.3 Dependencies

pip install -r requirements.txt

---

2.4 Kaolin

pip install kaolin==0.15.0
-f https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-2.0.0_cu118.html

---

2.5 PyTorch3D

pip install --no-index --no-cache-dir pytorch3d
-f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py39_cu118_pyt200/download.html

---

3. 🚀 Inference

3.1 Environment Variable

export BOP_DIR=/path/to/lm

---

3.2 LINEMOD

python run_linemod.py
--linemod_dir /path/to/lm

---

3.3 YCB-Video

python run_ycb_video.py
--ycbv_dir /path/to/YCB_Video

---

4. ⚠️ Troubleshooting

• Ensure CUDA matches PyTorch (recommended CUDA 11.8)
• RTX 4090+ → CUDA ≥ 12.1 preferred
• If build fails:
  • check Kaolin version
  • check PyTorch3D compatibility

---

5. 📊 Benchmark Results on LINEMOD

We evaluate on LINEMOD using ADD-0.1% metric.

Settings:

• RGB / RGB-D inputs
• 1-shot setting
• Hypotheses number N
• Real (*), Rendered (†)

---

5.1 Key Results Summary

Method	Year	Modality	Ref	Mean	Time
OnePose*	2022	RGB	200	63.6	66 ms
OnePose++*	2023	RGB	200	76.9	88 ms
FS6D*	2022	RGB-D	16	88.9	72 ms
SinRef-6D†	2025	RGB-D	1	90.2	-
Ours (N=12)*	2026	RGB-D	1	89.9	80 ms
Ours (N=78)*	2026	RGB-D	1	92.5	375 ms
Ours (N=12)†	2026	RGB-D	1	91.2	80 ms
Ours (N=78)†	2026	RGB-D	1	99.1	375 ms

---

5.2 Full LaTeX Table

\definecolor{highlightblue}{RGB}{235, 235, 255}

\begin{table*}[t] \centering \small \setlength{\tabcolsep}{1.8pt} \renewcommand{\arraystretch}{1.3}

\caption{LINEMOD (ADD-0.1%) comparison.} \label{tab:linemod}

\begin{tabular}{lcccccccccccccccccc} \toprule Method & Year & Mod. & Ref. & \multicolumn{13}{c}{Object ID} & Mean & Time \ \midrule

OnePose* & 2022 & RGB & 200 & ... \ OnePose++* & 2023 & RGB & 200 & ... \ FS6D* & 2022 & RGB-D & 16 & ... \ SinRef-6D† & 2025 & RGB-D & 1 & ... \

\midrule \rowcolor{highlightblue} Ours (N=12) & 2026 & RGB-D & 1 & ... \

\rowcolor{highlightblue} Ours (N=78) & 2026 & RGB-D & 1 & ... \

\bottomrule \end{tabular} \end{table*}