Generating Dynamic Human3.6M

June 1, 2022 ยท View on GitHub

Step 1: Download Human3.6M from the official website

Step 2: Preprocess Human3.6M into COCO format

We use the same COCO format for Human3.6M as PoseNet and I2L-MeshNet.
To generate the format, follow the instructions in the PoseNet repo:

  1. Run the matlab preprocessing script using the official Human3.6M SDK.
  2. Run h36m2coco.py: 
    python h36m2coco.py
  3. Download the SMPL parameters obtained using gradient-based optimization. Unzip the file into datasets/smpl_fit.

The resulting COCO-format Human3.6m dataset will have the following structure:

${GLAMR_ROOT}
|-- datasets
|   |-- H36M
|   |   |-- images
|   |   |-- annotations
|   |   |-- smpl_fit

Step 3: Generate Dynamic Human3.6M

Take the following steps in the root folder of this repo:

  1. Further process the COCO-format human3.6M dataset: 
    python preprocess/preprocess_h36m.py
  2. Generate Dynamic Human3.6M with occlusions: 
    python preprocess/preprocess_h36m_occluded.py

The resulting Dynamic Human3.6M dataset is stored in datasets/H36M/occluded_v2.