sapiens2-0.4b-pose-torch#

Sapiens2 0.4B (fastest) top-down whole-body pose estimation producing 308 keypoints across body, face, hands, and feet, pretrained on 1 billion human images. Requires Python 3.12+ and person box prompts via prompt_field.

Details

  • Model name: sapiens2-0.4b-pose-torch

  • Model source: https://huggingface.co/facebook/sapiens2-pose-0.4b

  • Model author: Rawal Khirodkar, et al.

  • Model license: Sapiens2 License

  • Model size: 1.59 GB

  • Exposes embeddings? no

  • Tags: keypoints, pose, human-centric, torch, transformer

Requirements

  • Packages: torch, torchvision, huggingface_hub, sapiens @ git+https://github.com/facebookresearch/sapiens2.git

  • CPU support

    • yes

  • GPU support

    • yes

Example usage

 1import fiftyone as fo
 2import fiftyone.zoo as foz
 3
 4dataset = foz.load_zoo_dataset(
 5    "coco-2017",
 6    split="validation",
 7    dataset_name=fo.get_default_dataset_name(),
 8    max_samples=50,
 9    shuffle=True,
10)
11
12model = foz.load_zoo_model("sapiens2-0.4b-pose-torch")
13
14dataset.apply_model(model, label_field="predictions")
15
16session = fo.launch_app(dataset)