FineGym-skeleton Dataset

License: CC BY 4.0

Overview

FineGym-skeleton is a human-skeleton action-recognition benchmark derived from FineGym. It combines temporally precise, fine-grained gymnastics annotations with 2D human-pose sequences extracted from the corresponding video frames. The current V2 release contains Gym99-skeleton-V2 and Gym288-skeleton-V2. The source RGB subaction clips are also available in the FineGym-RGB-subactions directory on Hugging Face.

The dataset supports research on:

• Fine-grained action recognition
• Temporally corrupted or incomplete action modeling
• Skeleton-based representation learning
• Physics-aware motion understanding

The dataset was introduced and used in FineTec: Fine-Grained Action Recognition Under Temporal Corruption via Skeleton Decomposition and Sequence Completion, published at AAAI 2026.

Update Log

• 2025-11-11: FineGym-skeleton V1 released with Gym288-skeleton.
• 2026-05-21: FineGym-skeleton V2 released with Gym99-skeleton and Gym288-skeleton. For V2, the first frame of every sample in both subsets was re-annotated to improve skeleton-extraction accuracy.

Annotation Pipeline

FineGym-skeleton was annotated using the Gym99 and Gym288 subsets of FineGym. For each RGB subaction clip, a bounding box was manually annotated on the first frame to identify the target gymnast. OSTrack was then used to track the target throughout the clip, and HRNet was applied to extract the frame-by-frame skeleton keypoints.

Released Data

FineGym-skeleton

Each subset includes predefined train and val splits that together cover all annotated samples without overlap.

Item	Gym99-skeleton	Gym288-skeleton
File	Gym99-skeleton-V2.pkl	Gym288-skeleton-V2.pkl
Fine-grained classes	99 (0–98)	288 (0–287)
Total annotations	34,803	38,935
Training samples	26,282	29,290
Validation samples	8,521	9,645
Total annotated frames	1,617,291	1,882,226
Frames per sample (min / mean / max)	2 / 46.47 / 725	2 / 48.34 / 725
People per sample	1	1
Keypoints per person	17	17
Coordinate / score dtype	float16 / float16	float16 / float16

The source videos occur at five resolutions in both subsets: (720, 1280), (1080, 1920), (480, 854), (360, 640), and (702, 1280), represented as (height, width).

Action Classes

Gym99-skeleton contains 99 classes and Gym288-skeleton contains 288 classes across four apparatuses:

• Floor Exercise (FX)
• Balance Beam (BB)
• Uneven Bars (UB)
• Vault — Women (VT)

Each class represents a specific gymnastics element. For the class definitions and mappings, refer to the FineGym project website and FineGym paper.

Pickle Structure

Each skeleton file is a Python dictionary with two top-level keys:

• split: a dictionary containing train and val, each a list of frame_dir strings.
• annotations: a list of dictionaries, one per action instance.

Each annotation has the following fields:

Key	Type	Shape / example	Description
frame_dir	str	"A0xAXXysHUo_002184_002237_0035_0036"	Unique action-clip identifier
label	int	93	Zero-based class label
img_shape	tuple	(720, 1280)	Source frame (height, width)
original_shape	tuple	(720, 1280)	Original source frame shape
total_frames	int	48	Number of frames in this skeleton sequence
keypoint	np.ndarray	(1, T, 17, 2)	(x, y) coordinates for 17 COCO-style keypoints
keypoint_score	np.ndarray	(1, T, 17)	Per-keypoint pose-estimation scores

The leading dimension of keypoint and keypoint_score is the number of people and is always 1 in these files. T equals total_frames and varies by sample.

Usage Example

import pickle

with open("Gym288-skeleton-V2.pkl", "rb") as file:
    data = pickle.load(file)

train_ids = data["split"]["train"]
val_ids = data["split"]["val"]

sample = data["annotations"][0]
print("Label:", sample["label"])
print("Frames:", sample["total_frames"])
print("Keypoints shape:", sample["keypoint"].shape)  # (1, T, 17, 2)

skeleton_sequence = sample["keypoint"][0]  # (T, 17, 2)

FineGym-RGB-subactions

FineGym-RGB-subactions contains the RGB subaction clips in MP4 format, organized into four parts:

Directory	MP4 files	Size
part1	10,000	2.04 GB
part2	10,000	2.20 GB
part3	10,000	2.19 GB
part4	9,092	1.95 GB
Total	39,092	8.39 GB

An example filename is 0LtLS9wROrk_E_000147_000152_A_0000_0005.mp4. The name preserves the source-video ID and the zero-padded temporal identifiers used by the FineGym event/action annotations.

Citation

We thank the authors of FineGym for their foundational work in fine-grained action recognition. If you use this dataset, please cite both FineTec and the original FineGym paper.

@article{shao2026finetec,
  title={FineTec: Fine-Grained Action Recognition Under Temporal Corruption via Skeleton Decomposition and Sequence Completion},
  volume={40},
  url={https://ojs.aaai.org/index.php/AAAI/article/view/37838},
  doi={10.1609/aaai.v40i11.37838},
  number={11},
  journal={Proceedings of the AAAI Conference on Artificial Intelligence},
  author={Shao, Dian and Shi, Mingfei and Liu, Like},
  year={2026},
  month={Mar.},
  pages={8842--8850}
}

License

This dataset is licensed under the Creative Commons Attribution 4.0 International License. You may share and adapt the material, including commercially, provided appropriate credit is given.

The underlying video content remains the property of its original sources (for example, YouTube uploaders). Please observe the original FineGym terms and applicable source-platform requirements when using RGB clips.

Acknowledgements

• Target gymnasts were tracked throughout the RGB subaction clips using OSTrack.
• Skeleton keypoints were extracted from the tracked gymnasts using HRNet.