KeypointNeRF: Generalizing Image-based Volumetric Avatars using Relative Spatial Encoding of Keypoints

Marko Mihajlovic · Aayush Bansal · Michael Zollhoefer . Siyu Tang · Shunsuke Saito

ECCV 2022

KeypointNeRF leverages human keypoints to instantly generate volumetric radiance representation from 2-3 input images without retraining or fine-tuning. It can represent human faces and full bodies.

News 🆕

• [2022/10/01] Combine ICON with our relative spatial keypoint encoding for fast and convenient monocular reconstruction, without requiring the expensive SMPL feature. More details are here.

Installation

Please install python dependencies specified in environment.yml:

conda env create -f environment.yml
conda activate KeypointNeRF

Data preparation

Please see DATA_PREP.md to setup the ZJU-MoCap dataset.

After this step the data directory follows the structure:

./data/zju_mocap
├── CoreView_313
├── CoreView_315
├── CoreView_377
├── CoreView_386
├── CoreView_387
├── CoreView_390
├── CoreView_392
├── CoreView_393
├── CoreView_394
└── CoreView_396

Train your own model on the ZJU dataset

Execute train.py script to train the model on the ZJU dataset.

python train.py --config ./configs/zju.json --data_root ./data/zju_mocap

After the training, the model checkpoint will be stored under ./EXPERIMENTS/zju/ckpts/last.ckpt, which is equivalent to the one provided here.

Evaluation

To extract render and evaluate images, execute:

python train.py --config ./configs/zju.json --data_root ./data/zju_mocap --run_val
python eval_zju.py --src_dir ./EXPERIMENTS/zju/images_v3

To visualize the dynamic results, execute:

python render_dynamic.py --config ./configs/zju.json --data_root ./data/zju_mocap --model_ckpt ./EXPERIMENTS/zju/ckpts/last.ckpt

(The first three views of an unseen subject are the input to KeypointNeRF; the last image is a rendered novel view)

We compare KeypointNeRF with recent state-of-the-art methods. The evaluation metric is SSIM and PSNR.

Models	PSNR ↑	SSIM ↑
pixelNeRF (Yu et al., CVPR'21)	23.17	86.93
PVA (Raj et al., CVPR'21)	23.15	86.63
NHP (Kwon et al., NeurIPS'21)	24.75	90.58
KeypointNeRF* (Mihajlovic et al., ECCV'22)	25.86	91.07

(*Note that results of KeypointNeRF are slightly higher compared to the numbers reported in the original paper due to training views not beeing shuffled during training.)

Reconstruction from a Single Image

Our relative spatial encoding can be used to reconstruct humans from a single image. As a example, we leverage ICON and replace its expensive SDF feature with our relative spatial encoding.

While it achieves comparable quality to ICON, it's much faster and more convinient to use (*displayed image taken from pinterest.com).

3D Human Reconstruction on CAPE

Models	Chamfer ↓ (cm)	P2S ↓ (cm)
PIFu (Saito et al., ICCV'19)	3.573	1.483
ICON (Xiu et al., CVPR'22)	1.424	1.351
KeypointICON (Mihajlovic et al., ECCV'22; Xiu et al., CVPR'22)	1.539	1.358

Check the benchmark here and more details here.

Publication

If you find our code or paper useful, please consider citing:

@inproceedings{Mihajlovic:ECCV2022,
  title = {{KeypointNeRF}: Generalizing Image-based Volumetric Avatars using Relative Spatial Encoding of Keypoints},
  author = {Mihajlovic, Marko and Bansal, Aayush and Zollhoefer, Michael and Tang, Siyu and Saito, Shunsuke},
  booktitle={European conference on computer vision},
  year={2022},
}

License

CC-BY-NC 4.0. See the LICENSE file.