SAM 3D Body: 1-10 Million Annotations Across 8 Human Pose Datasets

Meta released this collection of 1 to 10 million annotations in late 2024 to support 3D human body segmentation and pose estimation. It unifies labels from eight major computer vision sources including COCO, 3DPW, and EgoExo4D into a single framework for the Segment Anything Model (SAM).

Use Cases

• Training 3D human pose estimation models using unified labels from COCO and MPII
• Fine-tuning Segment Anything Models (SAM) for human body parts using SA1B annotations
• Analyzing egocentric human motion using the EgoExo4D and EgoHumans subsets

Strengths

• 1M to 10M annotation records
• Consolidates 8 major human pose datasets including COCO and 3DPW
• Includes egocentric perspective data from EgoExo4D and EgoHumans

Limitations

• Annotation-only repository requiring manual merging with original source images
• Inherent label noise from aggregating multiple heterogeneous source datasets with different original standards

Provenance

Source

Meta (Facebook) research team

Collection Method

Re-annotation and processing of existing public datasets

Freshness

Last updated November 2025.

Geography

Global