Skip to content

Latest commit

 

History

History
29 lines (25 loc) · 1.32 KB

README.md

File metadata and controls

29 lines (25 loc) · 1.32 KB
Raw

Human 3.6M Dataset (for epipolar transformer training)

  1. Download the H36M from the official website http://vision.imar.ro/human3.6m/description.php
  2. Process using https://github.com/CHUNYUWANG/H36M-Toolbox
  3. Undistort (optional)
python scripts/undistort_h36m.py
python scripts/undistort_h36m.py --anno ~/datasets/h36m/annot/h36m_train.pkl
~/datasets/h36m/undistortedimages

We can't share the dataset according to their license. Please visit http://vision.imar.ro/human3.6m/ to request access and contact the maintainer of the dataset.

To test without, please feed the model with a pair of images, the intrinsic and extrinsic matrices. Our visualization Ipython notebook provides a good example. https://github.com/yihui-he/epipolar-transformers/blob/master/READMD.md#1-epipolar-transformers-visualization

RHD Dataset (for 2D->3D lifting training)

Download using bash scripts

bash get_RHD.sh

Create soft link of datasets

ln -s ~/hand3d/RHD_published_v2/ .

Custom Dataset

If you are going to create your own dataset, please change the following places:

  • data/datasets/: dataloader
  • data/build.py: if clause in build_dataset
  • core/paths_catalog.py: DatasetCatalog.DATASETS and corresponding if clause in DatasetCatalog.get()