Skip to content

Latest commit

 

History

History
107 lines (95 loc) · 3.99 KB

README.md

File metadata and controls

107 lines (95 loc) · 3.99 KB
Raw

MonoPort Dataset

The folder structure is expected to be like this:

# |- MonoPortDataset
#     |- README.md
#     |- init_link.sh # link mixamo and renderppl data to ./data
#     |- bin/
#     |- lib/
#     |- scripts/
#     |- api/
#     |- data/
#         |- hdri/{*.exr}
#
#         |- mixamo/
#             |- {all, train, val}.txt;
#             |- actions  /<action>.fbx
#             |- skeletons/<action>/%06d.sk
#             |- clusters /kmeans/{all, train, val}_{k}.json
#
#         |- renderppl/
#             |- {all, train, val}.txt;
#             |- rigged    /<subject>_FBX/
#             |- tpose_objs/<subject>.obj
#             |- del_inside/<subject>/
#                {del_faces.npy, del_verts.npy}
#
#         |- pifu_orth/
#             |- {train, val}.txt;
#             |- <subject>/<action>/<frame>/
#                {mesh.obj, skeleton.txt, uv_render.png}
#                {calib/*.txt, render/*.png}
#
#     |- test_scripts/ # store experimental things.
#     |- test_data/ # store experimental things.

Dependence

  • tqdm
  • trimesh
  • blender:
cd ./MonoPortDataset
wget https://mirror.clarkson.edu/blender/release/Blender2.82/blender-2.82a-linux64.tar.xz -O ./bin/blender-2.82a-linux64.tar.xz
tar -xf ./bin/blender-2.82a-linux64.tar.xz -C ./bin/
  • tinyobjloader
  • vtkplotter (only for debug)

Note for scripts

init_link.sh: You need to setup the paths of where you store the data for renderppl and mixamo in this script. Then you can use this script to link the data into ./data/.

# under ./MonoPortDataset/
sh init_link.sh;

scripts/create_splits.sh: This script is how we did the <train/val> split for both renderppl and mixamo. It will create {all, train, val}.txt for both renderppl and mixamo.

# under ./MonoPortDataset/scripts/
bash create_splits.sh;

scripts/renderppl_tpose_objs.py: We use this script to export renderppl data to obj from fbx file. We then use the obj files to find those verts/faces inside the mesh (mouth, teeth, eyebows etc.)

# under ./MonoPortDataset/scripts/
../bin/blender-2.82a-linux64/blender --background --python renderppl_tpose_objs.py > /dev/null
# or using this line for multi blender instances processing
bash ./blender_multi_instances.sh renderppl_tpose_objs.py 20

scripts/renderppl_del_inside.py: We use this script to find those verts/faces inside the mesh (mouth, teeth, eyebows etc.) By default it runs using 8 threads.

# under ./MonoPortDataset/scripts/
python renderppl_del_inside.py;

scripts/mixamo_skeletons.py: We use this script to export skeletons from mixamo data.

# under ./MonoPortDataset/scripts/
../bin/blender-2.82a-linux64/blender --background --python mixamo_skeletons.py > /dev/null
# or using this line for multi blender instances processing
bash ./blender_multi_instances.sh mixamo_skeletons.py 20

scripts/mixamo_kmeans.py: As mixamo action data has severely unbalanced distribution, we perform kmeans here on mixamo data to do the clustering first.

# under ./MonoPortDataset/scripts/
python mixamo_kmeans.py --split all --klist 10 20 50 100 200 300 500 1000
python mixamo_kmeans.py --split train --klist 10 20 50 100 200 300 500 1000
python mixamo_kmeans.py --split val --klist 10 20 50 100

scripts/pifu_orth_splits.py: Here is how we create <train/val> splits for PIFu training. The logic here is for each renderppl subject, we randomly select a mixamo action cluster first, then randomly select a frame from it, and apply it to the renderppl subject. Note that this script is only for creating the lists, not actual rendered image.

# under ./MonoPortDataset/scripts/
python pifu_orth_splits.py

scripts/pifu_orth_render.py: Finally! we come to the render part!.

# under ./MonoPortDataset/scripts/
# use this line for multi blender instances processing
bash pifu_orth_render.sh 4

MonoPortDataset V2

../bin/blender-2.82a-linux64/2.82/python/bin/python3.7m inst