Synbody is the largest synthetic dynamic human gesture dataset. It includes a large scale of high-res human character models, motion sequences, varied scenes and rich modalities. Also, with the help of UnrealEngine, high-quality videos and annotations were generated.
SynBody builds dynamic scenes of multiple people in different styles of environments. The characters cover more than 700 human body gestures and models as data. Different clothing, body types, genders, and age groups are covered. It uses rich action types such as standing, walking, running, jumping, and dancing to drive the construction of human body models. It also provides the corresponding SMPL/SMPLX annotations.
This dataset complements the vacancy of large-scale dynamic human synthesis datasets in the academic field and will support various downstream tasks such as single-view human parametric model estimation, multi-view human parametric model estimation, and human/person presence detection and segmentation, and it aims to help promote virtual data Exploration of training methods.
The first batch of SynBody dataset synthesizes data from 5 camera perspectives in each dynamic scene. In addition to providing RGB images, it includes data annotations of 5 modalities, which can support a variety of human body-related research tasks. It significantly supports the exploration of the use of virtual human body data and the controlled number of characters, actions, and the angle of view of the camera. Plus, as the acquisition of real datasets is hampered by COVID-19, the SynBody dataset may contribute to the new paradigm of using synthetic datasets at scale for training models.
Large-scale dataset (>1,000,000 frames)
High-resolution human models (>750)
Diversified 3d scenes
5 camera views
6 data modalities