FineVideo-Phase2-3DPose — 3D Human Pose from MotionBERT

Overview

This dataset contains 3D human pose data lifted from 2D detections using MotionBERT, extracted from ~40K YouTube videos in the FineVideo dataset.

This is the output of Phase 2 (+ Phase 2.5 resampling) in the FineVideo-VLA pipeline. It contains raw 3D joint positions as NumPy arrays at 30fps, before any filtering, normalisation, or tokenisation.

Statistics

Metric	Value
Source videos	~40,000 from FineVideo
Videos processed	40,804
Total size	~259 GB (raw 3D) / ~67 GB (30fps resampled)
Frame rate	30 fps (resampled from native video fps)
Joints per frame	17 (H36M skeleton)

Pipeline Context

Phase	Description	Status
Phase 1	HRNet 2D pose detection (GPU)	Done
Phase 2	MotionBERT 2D-to-3D lifting (this dataset)	Done
Phase 2.5	Resample all videos to 30fps	Done
Phase 3	Kinematics: bone normalisation, root centering, smoothing	Done
Phase 4	YOLO person-detection cleaning	Done
Phase 5	Adaptive PCHIP per-joint tokenisation	Done
Phase 6	Merge agent tokens into multimodal dataset	Done
Phase 7	Flatten to Megatron-LM format	Done
Phase 8	Megatron-LM tokenization (.bin/.idx)	Done

Data Format

Each record contains 3D joint positions for one video as a NumPy array:

• Shape: (num_frames, 17, 3) — frames at 30fps, 17 joints, xyz coordinates
• Units: metres (MotionBERT output space)
• Coordinate system: camera-relative (not root-centred — root centering happens in Phase 3)

Joint order (H36M 17-joint skeleton)

Index	Joint	Index	Joint
0	pelvis (root)	9	nose
1	right hip	10	head top
2	right knee	11	left shoulder
3	right ankle	12	left elbow
4	left hip	13	left wrist
5	left knee	14	right shoulder
6	left ankle	15	right elbow
7	spine	16	right wrist
8	thorax

Processing Details

1. Phase 1 (HRNet): Ran HRNet with Faster R-CNN person detection to get 2D joint coordinates per frame
2. Phase 2 (MotionBERT): Lifted 2D poses to 3D using MotionBERT pretrained on Human3.6M, processed at native video fps
3. Phase 2.5 (Resample): Resampled from native video fps to uniform 30fps via linear interpolation, so poses align to the same time grid as video tokens (Seed2/Cosmos/AVC-LM)

Downstream Processing

For cleaned and normalised poses, see FineVideo-Phase4-YOLOPose which applies:

• Temporal smoothing (Butterworth filter)
• Bone length normalisation to canonical skeleton
• Root centering (pelvis at origin)
• Anti-teleportation filter
• YOLO person-presence cleaning

Related Resources

Resource	Description
EmpathicRobotics/FineVideo-Phase4-YOLOPose	Cleaned + normalised 3D poses (after Phase 3+4)
EmpathicRobotics/FineVideo-Phase5-AgentTokens	Merged multimodal dataset with tokenised pose + video tokens
EmpathicRobotics/FineVideo-Phase7-Flattened	Flat Megatron-LM JSONL (ready for pretraining)
EmpathicRobotics/tokenizer-vla-adaptive	HuggingFace tokenizer (144,215 vocab)

Usage

from datasets import load_dataset

ds = load_dataset("EmpathicRobotics/FineVideo-Phase2-3DPose", streaming=True)

for sample in ds["train"]:
    video_id = sample["video_id"]
    poses_3d = sample["poses"]  # (num_frames, 17, 3)
    print(f"Video: {video_id}, Frames: {len(poses_3d)}")
    break

Citation

Part of the FineVideo-VLA project. If you use this data, please cite:

@misc{Farré2024FineVideo,
  title={FineVideo},
  author={Farré, Miquel and Marafioti, Andi and Tunstall, Lewis and Von Werra, Leandro and Wolf, Thomas},
  year={2024},
  howpublished={\url{https://huggingface.co/datasets/HuggingFaceFV/finevideo}},
}

License

Apache 2.0