Skip to content

Latest commit

 

History

History
45 lines (32 loc) · 3.41 KB

data.md

File metadata and controls

45 lines (32 loc) · 3.41 KB
Raw

Despription

We assemble a quite large dataset (named as WILD dataset in the paper) for training and testing by extracting from 5 human datasets, including Leeds Sport Pose dataset (LSP) and its extension dataset (LSPet), MPII human pose database (MPII), Common Objects in Context dataset (COCO), and Human3.6M dataset (H36M).

We filter out images with incomplete human or low-quality silhouette, and re-arrange the data in the order of "LSP - LSPET - MPII - COCO - H36M". A data log will be produced during the generation. In each line, the log is written as:

[src_dataset]  [src_num]  [data_type]  [new_num]

  • [src_dataset]: enum {LSP, LSPET, MPII, COCO, H36M}.

  • [src_num]: ID number in the source dataset.

  • [data_type]: enum {TRAIN (select as training data), TEST (select as testing data), BAN (not selected)}

  • [new_num]: ID number in the WILD dataset.

This will help trace the source of each tuple of the data.

HMR configure

We use HMR to predict the initial shape for our method. In order to speed up the training, we pre-compute the HMR results and save them in the dataset. The HMR results include predicted keypoint position, SMPL mesh, rendered silhouette and other related attributes. HMR repository is required during the data processing. Please follow the instructions here to setup the enviroment for HMR, and make sure the basic demo in HMR works. Then set the HMR location to HMR_PATH in "/conf.ini".

Data acquire

Download and extract the source data, then set the path in the [DATA] part of "/conf.ini" as following:

  • LSP. Acquire the data from http://sam.johnson.io/research/lsp.html by clicking "Download (ZIP file)". Extract the data and set the extracting path to lsp_path.

  • LSPet. Acquire the data from http://sam.johnson.io/research/lspet.html by clicking "Dataset (ZIP file)". Extract the data and set the extracting path to lspet_path.

  • UP (contains MPII Human Shape data). Acquire the data from http://files.is.tuebingen.mpg.de/classner/up/ by clicking "UPi-S1h (26,294 images, 44.3Gb)". Extract the data and set the extracting path to upi_path.

  • H36M. Acquire the data from http://vision.imar.ro/human3.6m/description.php. Register and apply for the access to the data. The selectedd data are listed in "h36m_list.txt". Extract the data and set the extracting path to h36m_path. The joints in H36M dataset are saved in cdf files. As the cdf accessor in python is quite difficult to install, we use matlab to tansform all the .cdf files to .mat files, which are easier to access in python. The transforming script is saved in h36m_cdf2mat.m.

  • COCO. Visit http://cocodataset.org/#download to download the COCO API, then set the API path to coco_api_path. COCO data is stored online, so the data will be downloaded through the API while processing. Download the data list (json file) for densepose as following and set the path to coco_list_path:

cd $certain path$
wget https://dl.fbaipublicfiles.com/densepose/densepose_coco_2014_train.json

Note: Please follow the license agreements of the above-mentioned datasets.

Run the processing

Finally, set the path to save the dataset in tgt_path and run the processing:

cd datasets
python proc_all_data.py

The dataset requires 26 GB space, and the processing will last for roughly 11 hours. We provide the download links of testing set in Demo part.