In this section, you will learn how to utilize the results of motion capture to achieve 3D reconstruction of the human body and free viewpoint rendering. Our approach primarily relies on neuralbody, but similar methods such as aninerf can also be employed.
Try the code
Preprocess
We will render the previously generated SMPL results into instance maps and vertices.
data=<path/to/boxing>
mkdir -p ${data}/output-smpl-3d
cp -r output/boxing/smpl ${data}/output-smpl-3d
# convert the SMPL parameter to vertices
python3 apps/postprocess/write_vertices.py ${data}/output-smpl-3d/smpl ${data}/output-smpl-3d/vertices --cfg_model config/model/smpl.yml --mode vertices
# render the instance map
python3 apps/postprocess/render.py ${data} --exp output-smpl-3d --mode instance-d0.05 --ranges 0 200 1 --model config/model/smpl.yml
Check the dataset can be loaded rightly:
python3 apps/neuralbody/check_trainset.py --cfg config/neuralbody/dataset/multiview/lightstage.yml --opts split train data_share_args.root ${data}