Skip to content

Latest commit

 

History

History
48 lines (34 loc) · 2.76 KB

2D_pose.md

File metadata and controls

48 lines (34 loc) · 2.76 KB

2D Human Pose Estimation Datasets

For 2D human pose estimation, we adopt the COCO format and use the datasets provided by mmpose.

COCO

Please follow the dataset preparation steps in mmpose to prepare the COCO datasets.

AIC

Please follow the dataset preparation steps in mmpose to prepare the AIC datasets.

Human3.6m_pose

Please follow the dataset preparation steps in mmpose to prepare the Human3.6m_pose datasets. We provide the processed 2D pose annotation json in the 2dpose.

After copying the 2D pose annotation json to the corresponding folder, the folder structure should look like:

- h36m
    - annotation_body2d
        - h36m_coco_train.json
        - h36m_coco_test.json
    - annotation_body3d
    - images

posetrack

Please follow the dataset preparation steps in mmpose to prepare the posetrack datasets.

jrdb2022

Please log in the original website to download the dataset.

MHP v2

Please download the dataset from the official website.

mpi_inf_3dhp

Please download the dataset from the official website.

3dpw

Please download the dataset from the official website. The json file can be found in the 2dpose folder.

AIST++

We project 3D poses in AIST++ into 2D poses. Please download the dataset from the official website. As it is a video dataset, we do not use all the frames for training, the training json of sampled framed cound be found in the 2dpose folder.

Halpe (not used in Hulk, but used in UniHCP)

Please follow the dataset preparation steps in mmpose to prepare the Halpe datasets. We remove the the joints of hands and face. The filtered json file can be found in the 2dpose folder.