alecccdd/celeb-fbi-pose-estimation

Celeb-FBI Pose Estimation Dataset

A derivative of the Celeb-FBI dataset enriched with 3D human pose and mesh recovery annotations generated using Meta's SAM 3D Body model.

Dataset Description

This dataset extends the original Celeb-FBI celebrity full-body image dataset with comprehensive 3D pose estimation outputs. Each sample includes predicted 3D keypoints, 2D keypoints, mesh vertices, body/hand pose parameters, and shape parameters extracted using the facebook/sam-3d-body-vith model.

The dataset retains the original biometric annotations (height, weight, age, gender) from Celeb-FBI, enabling joint research on pose estimation and human attribute prediction.

Dataset Structure

DatasetDict({
    train: Dataset({
        features: ['id', 'bbox', 'focal_length', 'pred_keypoints_3d', 'pred_keypoints_2d', 
                   'pred_vertices', 'pred_cam_t', 'pred_pose_raw', 'global_rot', 
                   'body_pose_params', 'hand_pose_params', 'scale_params', 'shape_params', 
                   'expr_params', 'mask', 'pred_joint_coords', 'pred_global_rots', 
                   'lhand_bbox', 'rhand_bbox', 'height', 'weight', 'gender', 'age'],
        num_rows: 5499
    })
    test: Dataset({
        features: [...],
        num_rows: 612
    })
})

Features

Pose Estimation Outputs (SAM 3D Body)

Feature	Type	Description
bbox	sequence[float]	Detected person bounding box
focal_length	float	Estimated camera focal length
pred_keypoints_3d	sequence[sequence[float]]	Predicted 3D keypoint coordinates
pred_keypoints_2d	sequence[sequence[float]]	Predicted 2D keypoint coordinates
pred_vertices	sequence[sequence[float]]	Predicted mesh vertices (MHR format)
pred_cam_t	sequence[float]	Predicted camera translation
pred_pose_raw	sequence[float]	Raw pose parameters
global_rot	sequence[float]	Global rotation parameters
body_pose_params	sequence[float]	Body pose parameters
hand_pose_params	sequence[float]	Hand pose parameters
scale_params	sequence[float]	Scale parameters
shape_params	sequence[float]	Body shape parameters
expr_params	sequence[float]	Expression parameters
mask	Image	Segmentation mask
pred_joint_coords	sequence[sequence[float]]	Predicted joint coordinates
pred_global_rots	sequence[sequence[float]]	Predicted global rotations per joint
lhand_bbox	sequence[float]	Left hand bounding box
rhand_bbox	sequence[float]	Right hand bounding box

Biometric Annotations (from Celeb-FBI)

Feature	Type	Description
id	string	Unique identifier (matches original Celeb-FBI)
height	float	Height in centimeters (-1 if missing/invalid)
weight	float	Weight in kilograms (-1 if missing/invalid)
gender	int	0 = Male, 1 = Female
age	int	Age in years (-1 if missing/invalid)

Sample Count

Split	Original Celeb-FBI	This Dataset	Notes
Train	6,487	5,499	Single-person images only
Test	721	612	Single-person images only
Total	7,208	6,111	~85% retained

Note: Images where the SAM 3D Body model detected ≠1 person have been excluded to ensure clean, unambiguous pose annotations.

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("alecccdd/celeb-fbi-pose-estimation")

# Access training data
train_data = dataset["train"]

# Example: iterate over samples
for sample in train_data:
    # Pose estimation outputs
    keypoints_3d = sample["pred_keypoints_3d"]
    keypoints_2d = sample["pred_keypoints_2d"]
    body_pose = sample["body_pose_params"]
    shape = sample["shape_params"]
    
    # Original biometric data
    height = sample["height"]  # in cm, -1 if missing
    weight = sample["weight"]  # in kg, -1 if missing
    gender = sample["gender"]  # 0=male, 1=female
    age = sample["age"]        # -1 if missing

# Filter samples with valid height annotations
valid_height = train_data.filter(lambda x: x["height"] != -1)

# Get samples with both pose and complete biometric data
complete_samples = train_data.filter(
    lambda x: x["height"] != -1 and x["weight"] != -1 and x["age"] != -1
)

Accessing Original Images

The original images are available in the base dataset:

from datasets import load_dataset

# Load both datasets
pose_data = load_dataset("alecccdd/celeb-fbi-pose-estimation")
image_data = load_dataset("alecccdd/celeb-fbi")

# Create a lookup by ID
image_lookup = {sample["id"]: sample["image"] for sample in image_data["train"]}

# Access image for a pose sample
pose_sample = pose_data["train"][0]
original_image = image_lookup.get(pose_sample["id"])

Processing Pipeline

This dataset was created by:

1. Loading all images from alecccdd/celeb-fbi
2. Running inference with facebook/sam-3d-body-vith on each image
3. Filtering to retain only images where exactly one person was detected
4. Extracting and storing all pose estimation outputs alongside original annotations

Intended Uses

• Pose-conditioned attribute estimation: Using 3D pose features to improve height, weight, or age prediction
• Multi-task learning: Joint pose estimation and biometric prediction
• Pose analysis: Studying pose distributions in celebrity photography
• Benchmarking: Evaluating pose estimation models on standing, front-facing poses
• Shape-from-pose research: Investigating relationships between body shape parameters and physical attributes

Limitations

• Single-person constraint: Multi-person images have been removed
• Celebrity bias: Images may not represent general population poses and body types
• Pseudo-ground-truth: Pose annotations are model predictions, not manual annotations
• Standing poses: Original dataset focuses on front-facing, standing positions
• Inherited noise: Some biometric annotations from Celeb-FBI contain errors

Ethical Considerations

• This dataset uses publicly available images of celebrities
• Pose estimation can be used for surveillance; users should consider privacy implications
• Body shape parameters may encode sensitive attributes
• Please use responsibly and in accordance with applicable laws

Citation

If you use this dataset, please cite both the SAM 3D Body model and the original Celeb-FBI paper:

@article{yang2025sam3dbody,
  title={SAM 3D Body: Robust Full-Body Human Mesh Recovery},
  author={Yang, Xitong and Kukreja, Devansh and Pinkus, Don and Sagar, Anushka and Fan, Taosha and Park, Jinhyung and Shin, Soyong and Cao, Jinkun and Liu, Jiawei and Ugrinovic, Nicolas and Feiszli, Matt and Malik, Jitendra and Dollar, Piotr and Kitani, Kris},
  journal={arXiv preprint},
  year={2025}
}

@misc{debnath2024celebfbibenchmarkdatasethuman,
  title={Celeb-FBI: A Benchmark Dataset on Human Full Body Images and Age, Gender, Height and Weight Estimation using Deep Learning Approach}, 
  author={Pronay Debnath and Usafa Akther Rifa and Busra Kamal Rafa and Ali Haider Talukder Akib and Md. Aminur Rahman},
  year={2024},
  eprint={2407.03486},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2407.03486}, 
}

Related Resources

• Base Dataset: alecccdd/celeb-fbi
• SAM 3D Body Model: facebook/sam-3d-body-vith
• SAM 3D Body Code: github.com/facebookresearch/sam-3d-body
• SAM 3D Body Paper: arXiv
• Original Celeb-FBI Paper: arXiv:2407.03486