About SynPose
SynPose is a large, densely labeled synthetic dataset specifically designed for crowded human pose estimation in classroom and meeting scenarios. Our dataset is constructed by exploiting Grand Theft Auto V (GTA V), a highly photorealistic open-world game developed by Rockstar North. Below is a demo video of the dataset construction.
For more information about the properties of the dataset please click into the tabs above.
Evaluation visualization
We propose a novel Classroom-style Transfer Generative Adversarial Network (CTGAN) to reduce the domain gap by transferring synthetic images to realistic classroom style. Thus, these translated synthetic images are used to train a pose estimation model and test on our collected real-classroom images. Here are more visualization results:
CTGAN evaluation
We adapt the technique of CycleGAN to learn a translation mapping between SynPose and real-world classroom images. Moreover, we add an id-loss to avoid the color changes and a mask-loss to
remove the artifacts in translated images.
Click below to see more visualizations about comparison of different constraints in CTGAN.
See more..
Pose estimation evaluation
We train the pose estimation model on COCO, Crowdpose and our SynPose subsets then test on real-world classroom images.
Results show that our dataset can help improve the performance of the model in crowded classroom scenarios.
Click below to see more visualizations about qualitative results of pose estimation evaluation on real-classroom images.
See more..
Ethical Considerations
People-centric datasets pose ethical challenges. Our dataset were created with careful attention to ethical questions, which we encountered throughout our work. Importantly, we are very cautious of automatic annotation procedure of SynPose datasets towards the social and ethical implications. The characters in the dataset were randomly selected without any personal bias. Characters are only labeled with poses and bounding-boxes without any labeling about appearance (such as gender, skin, etc.). Access to our dataset will be provided for research purposes only and with restrictions on redistribution.