Foundation for Human Vision Models

Sapiens is a foundation for human vision models developed by Facebook Research, aimed at enhancing human-centric vision tasks such as 2D pose estimation, part segmentation, depth estimation, and normal estimation. The models are pretrained on a dataset of 300 million human images and optimized for high-resolution feature extraction at 1024 x 1024 pixels. Users can get started by cloning the repository and setting the SAPIENS_ROOT path. There are two installation options: a lite version for inference-only use and a full installation for replicating the complete training setup. Sapiens supports various tasks including image encoding, pose estimation, body part segmentation, depth estimation, and surface normal estimation. Guides for finetuning the models are also provided, particularly for surface normal estimation. The project acknowledges contributions from OpenMMLab and encourages users to report issues for support. It is licensed under specific terms outlined in the repository, and users are encouraged to cite the project in their research using the provided BibTeX entry.

- Sapiens is designed for human-centric vision tasks and pretrained on 300 million images.

- The models operate at a resolution of 1024 x 1024 pixels.

- Users can choose between a lite installation for inference or a full installation for training.

- The repository includes guides for various tasks and finetuning models.

- Contributions from OpenMMLab are acknowledged, and users can cite the project in research.