README.md

---
license: mit
task_categories:
- graphics
tags:
- 3d
- face
- animation
- retargeting
- siggraph
size_categories:
- 1K<n<10K
---

# NFR sample training data

Sample data pack for [NFR (Neural Face Rigging)](https://github.com/dafei-qin/NFR_pytorch),
SIGGRAPH 2023. It covers the `ICT_single_coarse` dataset (40 identities) and is
meant to get the training pipeline running without downloading the full ~800GB
set used in the paper.

## What is in here

Only the files that **cannot be regenerated** are shipped (3.0GB):

```
data/ICT_single_coarse/{train,val,test}/*_processed_matrix.pkl   # 40 x 75MB, deformation gradients
data/ICT_single_coarse/*_neutral.obj                             # 40 neutral meshes
data/ICT_single_coarse/iden_exp_weights_single.npz               # identity / expression coefficients
data/ICT_single_coarse/ICT_single_exp_weights.npy                # per-expression weights
data/ICT_single_coarse/gradients_std.npy
data/ICT_live_100/gradients_std.npy                              # referenced by std_file
data/MF_all_v5/img_stat.pkl                                      # referenced by img_file
```

Each `_processed_matrix.pkl` holds 266 expressions of one identity: `verts`
(266, 3694, 3) and `gradients` (266, 7007, 9).

## Usage

Download and unpack at the repository root:

```python
from huggingface_hub import snapshot_download
snapshot_download(repo_id="HKU-CGVU/NFR-sample-data", repo_type="dataset",
                  local_dir=".", allow_patterns="data/*")
```

`allow_patterns="data/*"` keeps this dataset card from landing on top of the
repository's own `readme.md` on case-insensitive filesystems (macOS, Windows).

Then regenerate the derived tensors (renders, DiffusionNet gradients and
spectral decompositions):

```bash
python dataset_preprocess.py \
    --data_dir ./data/ICT_single_coarse \
    --save_data_dir ./data/ICT_single_coarse \
    -pi -pg -pd --num_threads 8
```

This adds about 32GB and takes roughly 30 minutes. Then train:

```bash
python FWH_run.py -c config/train_single.yaml
```

Training this configuration needs about 24GB of RAM and a single GPU. See the
Training section of the repository readme for details.

## Source

Derived from [ICT-FaceKit](https://github.com/ICT-VGL/ICT-FaceKit). Please also
follow the ICT-FaceKit license terms.

## Citation

```bibtex
@inproceedings{qin2023NFR,
  title={Neural Face Rigging for Animating and Retargeting Facial Meshes in the Wild},
  author={Qin, Dafei and Saito, Jun and Aigerman, Noam and Groueix, Thibault and Komura, Taku},
  booktitle={SIGGRAPH 2023 Conference Papers},
  year={2023}
}
```