H3WB: Human3.6M 3D WholeBody Dataset

H3WB Dataset Overview

Dataset Description

• Name: H3WB: Human3.6M 3D WholeBody Dataset and Benchmark
• Type: Large‑scale 3D full‑body pose estimation dataset
• Extended From: Human3.6M dataset
• Content: 133 full‑body keypoints (17 body, 6 foot, 68 face, 42 hand) annotated on 100,000 images
• Keypoint Layout: Same as COCO‑WholeBody dataset

Dataset Download

• Images: Available from the Human3.6M official website
• Annotations: Downloadable from the specified link, default location datasets/json/
• Formatted Data: A common format download is provided together with data‑preparation scripts, recommended for experiments

Annotation Format

• Each JSON file contains sample ID, image path, bounding box, 2D keypoints, and 3D keypoint information

Pre‑trained Models

• The dataset includes pre‑trained models used to generate the data; implementations are located in the models/ directory

Tasks

• 2D → 3D: From complete 2D full‑body to 3D full‑body
• I2D → 3D: From incomplete 2D full‑body to 3D full‑body
• RGB → 3D: From image to 3D full‑body prediction

Evaluation

• Test Set: Published, includes 2D→3D and I2D→3D test sets, each containing 10,000 {image, 2D coordinates, 3D coordinates} triples
• Validation: No dedicated validation set; 5‑fold cross‑validation results are encouraged

Citation Information

• Please follow the BibTeX entry in the README for proper citation