tiangexiang / wild2avatar

Official implementation for 'Wild2Avatar: Rendering Humans Behind Occlusions'

Wild2Avatar: Rendering Humans Behind Occlusions

Paper: https://arxiv.org/pdf/2401.00431.pdf

UNDER ACTIVE CONSTRUCTION 🚧

Dataset

We sampled 6 sequences from the origianl OcMotion dataset to be used in our experiments. Each sequence has 100 frames with precomputed segmentation masks and ground truth SMPL parameters.
The pre-proccesed sequences can be found here. Please follow the dataloader provided in OccNeRF to load the downloaded sequences.
If you find the preprocessed sequences usefull, please consider to cite CHOMP.

Citation

If you find this repo useful in your work or research, please cite:

@article{xiang2023wild2avatar,
    title={Wild2Avatar: Rendering Humans Behind Occlusions},
    author={Tiange Xiang and Adam Sun and Scott Delp and Kazuki Kozuka and Li Fei-Fei and Ehsan Adeli},
    journal={arXiv},
    year={2023}
	}

wild2avatar's Issues

Code release.

Hi @tiangexiang , I want to know if there is a plan for code release?

Training time

Hi,I'd like to know how long does it take to train the model?

Interpretations of One Ablation

Hi @tiangexiang et al.,

Thx for ur great constant contribution in this field. May I ask in Sec. 4.6 & Fig. 8 1st row, it states:

... by inheriting the two-stage parameterization but optimizing the rendering of the foreground human on the visible parts only [46] while maintaining the same optimization objectives.

if this is the case, then why it will lead to bad rendering results in Fig. 8? I am a bit confused and may miss sth. Is that becuz there are no sufficient frames where the problematic part is visible? Many thx in advance for any help. Rgds,