README.md

metadata

license: apache-2.0
task_categories:
  - video-classification
  - robotics
tags:
  - 3d-pose
  - human-pose-estimation
  - motion
  - finevideo
  - vla
language:
  - en
size_categories:
  - 10M<n<100M

FineVideo Phase 4 — YOLO-Cleaned 3D Human Pose (30fps)

Overview

This dataset contains YOLO-cleaned, bone-normalised 3D human pose data extracted from ~40K YouTube videos in the FineVideo dataset. It is the output of Phase 4 in the FineVideo-VLA pipeline and serves as input to Phase 5 (adaptive PCHIP tokenisation for LLM pretraining).

Use this dataset if you need raw 3D joint positions (floats in metres, not tokenised). For tokenised versions, see the related datasets below.

Statistics

Metric	Value
Source videos	~40,000 from FineVideo
Videos after cleaning	40,195
Total size	~107 GB (uncompressed JSONL)
Frame rate	30 fps (resampled from native video fps)
Joints per frame	17 (H36M skeleton)
Frames per window	8 (~0.267 seconds)

Pipeline Context

This dataset is part of a multi-phase pipeline that produces the FineVideo-VLA multimodal pretraining dataset:

Phase	Description	Status
Phase 1	HRNet 2D pose detection (GPU)	Done
Phase 2	MotionBERT 2D→3D lifting (GPU)	Done
Phase 2.5	Resample all videos to 30fps	Done
Phase 3	Kinematics: bone normalisation, root centering, smoothing, hallucination filtering	Done
Phase 4	YOLO person-detection cleaning (this dataset)	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 is a JSON line:

{
  "video_id": "abc123XYZ",
  "window_id": 320,
  "states": [[[x, y, z], ...17 joints...], ...8 frames...]
}

Field	Type	Description
video_id	string	YouTube video ID
window_id	int	Absolute frame index of the first frame in this window
states	float[8][17][3]	3D joint positions in metres

Timestamp

Absolute timestamp from video start: window_id / 30.0 seconds.

Each window covers 8 frames = 8/30 = 0.267 seconds.

Joint coordinates

• Root-centred: pelvis (joint 0) is always at origin [0, 0, 0]
• Bone-normalised: skeleton retargeted to canonical bone lengths
• Smoothed: temporal smoothing + anti-teleportation filter applied in Phase 3
• Coordinate range: typically +/-0.5m, max +/-2.0m

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

Skeleton connectivity

        head_top (10)
            |
          nose (9)
            |
        thorax (8)
       /    |    \
  l_sh(11) spine(7) r_sh(14)
   |        |        |
 l_el(12) pelvis(0) r_el(15)
   |      /    \      |
 l_wr(13) l_hip r_hip r_wr(16)
         (4)    (1)
          |      |
        l_kn   r_kn
         (5)    (2)
          |      |
        l_an   r_an
         (6)    (3)

Window structure

• Each window = 8 consecutive frames at 30fps (~0.267 seconds)
• window_id = absolute frame index (always a multiple of 8 after stride filtering)
• Absolute timestamp: window_id / 30.0 seconds from video start

YOLO cleaning (Phase 4)

Windows are dropped if >= 4 of 8 frames have no person detected by YOLOv8 (confidence >= 0.75). This removes windows where the subject is off-screen, occluded, or in a scene transition.

Some windows may still contain null/NaN values for individual joints where the pose estimator failed — downstream consumers should check for this.

Related Resources

Resource	Description
EmpathicRobotics/FineVideo-Phase5-AgentTokens	Merged multimodal dataset with tokenised pose + video tokens (hierarchical, full metadata)
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
import numpy as np

ds = load_dataset("EmpathicRobotics/FineVideo-Phase4-YOLOPose", streaming=True)

for sample in ds["train"]:
    video_id = sample["video_id"]
    window_id = sample["window_id"]
    states = np.array(sample["states"])  # (8, 17, 3)
    timestamp = window_id / 30.0         # seconds from video start

    print(f"Video: {video_id}")
    print(f"Window: {window_id} ({timestamp:.3f}s)")
    print(f"Pelvis (frame 0): {states[0, 0]}")  # always [0, 0, 0]
    print(f"Right wrist (frame 0): {states[0, 16]}")
    break

Citation

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

@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