Our dataset focuses on 3D human motion, we capturing actions where humans interact with chairs, encompassing both individual chair interactions and more complex scenes. Moreover, our dataset includes novel chair designs and novel sitting postures.
The file structure of our dataset is as follows:
/CHAIR:3D Human-Interaction Dataset
├── dataset
| ├── motion file of a model
| | ├──motion of several actions in succession
| | | ├──AMC
| | | | transl.npy
| | | | global_orient.npy
| | | | body_pose.npy
| | | | betas.npy
| | | | ...
| | | |
| | | ├──C3D
| | | obj_id.npy
| | | obj_transl.npy
| | | obj_orient.npy
| | | ...
| | |
| ├── lable
|
├── source_files
| ├── chairs_img
| ├── chairs_obj
| ├── model_fbx
| ├── ...
|
├── videos
| ├── video of a model
| ├── ...
|Dataset Contents: Our dataset comprises human motion data files organized into separate folders for each model's motion. Each motion is placed within its own subfolder. The motion files include SMPL-X parameters, original ASF/AMC files, and information regarding object motion.
Source Files: The "source_files" directory contains files related to object models and models themselves. It includes .obj format object model files and .fbx format model files.
Videos: The "videos" directory contains demonstration videos of human motions, with subdirectories named in the same way as the dataset's subdirectories.
Output Formats: The dataset provides motion data in various formats, including C3D, ASF/AMC, BVH, SMPL-X , information regarding object motion, and demonstration videos.
Data Volume: The dataset comprises data from six models, featuring approximately 280 motion sequences. There are interactions with 32 different objects, with over 20 objects exhibiting motion during the interactions. Each motion sequence consists of six consecutive actions. The duration of each motion sequence ranges from 30 to 45 seconds, with a frame rate of 60 FPS. The total duration of the dataset is approximately 2.8 hours.
We recommend using amc2bvh, which is open-source and free. Although amc2bvh does not consider scaling, we can set the global scale when importing .bvh into Blender.
In addition, we modified the first frame of the .amc file to match the T-pose (initial body_pose) of smpl-x.
To create a SMPL-X model in blender, we recommend the SMPL-Blender-Addon. Addon, you can manually set the model's betas parameters, height and weight. see the Tutorial.
The root bone of the model created with the SMPL-X Blender Add-on named 'root', which can be used in Blender to represent global displacement.
However, it is not necessary for describing the pose of the SMPL-X model. In fact, the bone hierarchy with 'pelvis' as the root bone is more similar to the structure of ASF skeletons, which is crucial for the retargeting. So, we removed the 'root' bone of the model, making 'pelvis' the root bone instead.
Then, we used the Rokoko Studio live for Blender for retargeting. The bone_mapping is provided.
The betas and body_pose per frame can be directly exported, and the SMPL_Blender_add-on code can be used directly in Blender Python scripts. Additionally, after obtaining betas, body_pose, joint_position and mesh_vertices, you can calculate global_transl and global_orient. (The process is similar to solving a 3D rigid body transformation, if any unclear about 3D rigid body transformation, reference: https://www.researchgate.net/publication/226330949)
All the scripts of data processing are located in the './scripts'.