---
license: mit
pipeline_tag: other
tags:
- semantic-correspondence
- 3d
- keypoints
---

# SemanticTransfer Pretrained Weights

Pretrained checkpoints for [Semantic Correspondence via 2D-3D-2D Cycle](https://arxiv.org/abs/2004.09061) (You et al., 2020), which predicts semantic correspondences by lifting 2D images to 3D and projecting corresponding 3D models back to 2D with semantic labels.

Code: https://github.com/qq456cvb/SemanticTransfer

## Contents

| File | Network | Size |
|---|---|---|
| `marrnet1.pt` | MarrNet-1: 2.5D sketch (depth/normal/silhouette) estimation | 123 MB |
| `shapehd.pt` | ShapeHD: 3D shape completion from 2.5D sketches | 277 MB |
| `best.pt` | Viewpoint (azimuth/elevation) estimation network | 435 MB |

## Usage

Download into the repository's `weights/` folder:

```bash
hf download qq456cvb/SemanticTransfer marrnet1.pt shapehd.pt best.pt --local-dir weights
```

Then run the demo:

```bash
python demo.py
```

## Citation

```bibtex
@article{you2020semantic,
  title={Semantic Correspondence via 2D-3D-2D Cycle},
  author={You, Yang and Li, Chengkun and Lou, Yujing and Cheng, Zhoujun and Ma, Lizhuang and Lu, Cewu and Wang, Weiming},
  journal={arXiv preprint arXiv:2004.09061},
  year={2020}
}
```