Gym288-skeleton Dataset

License: CC-BY-4.0

Overview

The Gym288-skeleton dataset is a human skeleton-based action recognition benchmark derived from the Gym288 subset of the FineGym dataset. It provides temporally precise, fine-grained annotations of gymnastic actions along with 2D human pose sequences extracted from original video frames.

This dataset is designed to support research in:

• 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 the paper "FineTec: Fine-grained Action Recognition under Temporal Corruption", which has been accepted to AAAI 2026. In this work, the dataset serves as the primary evaluation benchmark for recognizing fine-grained actions from temporally corrupted skeleton sequences.

Dataset Structure

The dataset is distributed as a single Python dictionary with two top-level keys: split and annotations.

Top-Level Keys

• split: Dictionary containing train/test splits.

  • train: List of 28,739 sample IDs (strings)
  • test: List of 9,484 sample IDs (strings)

• annotations: List of 38,223 dictionaries, each representing one action instance with the following fields:

Key	Type	Shape / Example	Description
frame_dir	str	"A0xAXXysHUo_002184_002237_0035_0036"	Unique identifier for the action clip
label	int	268	Class label (0–287, corresponding to 288 fine-grained gymnastic elements)
img_shape	tuple	(720, 1280)	Height and width of original video frames
original_shape	tuple	(720, 1280)	Same as img_shape (for compatibility)
total_frames	int	48	Number of frames in the action sequence
keypoint	np.ndarray (float16)	(1, T, 17, 2)	2D joint coordinates (x, y) for 17 COCO-style keypoints over T frames
keypoint_score	np.ndarray (float16)	(1, T, 17)	Confidence scores for each keypoint
kp_wo_gt	np.ndarray (float32)	(T, 17, 3)	Placeholder array (all zeros); originally intended for corrupted/noisy poses without ground truth
kp_w_gt	np.ndarray (float32)	(T, 17, 3)	Ground-truth 2D poses with confidence as third channel (x, y, score)

Note: The first dimension (1) in keypoint and keypoint_score corresponds to the number of persons (always 1 in this dataset).

Action Classes

The dataset contains 288 distinct gymnastic elements across four apparatuses:

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

Each class represents a highly specific movement (e.g., "Switch leap with 0.5 turn", "Clear hip circle backward with 1 turn to handstand"), reflecting the fine-grained nature of competitive gymnastics scoring.

For the full list of class names and mappings, please refer to the website and paper of FineGym.

Usage Example

import numpy as np

# Load the dataset (e.g., using pickle or torch.load)
with open("gym288_skeleton.pkl", "rb") as f:
    data = pickle.load(f)

# Access training samples
train_ids = data["split"]["train"]  # list of strings

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

# Extract skeleton sequence for model input
skeleton_seq = sample["keypoint"][0]  # (T, 17, 2)

Citation

If you use this dataset in your research, please cite both the FineTec and FineGym papers. FineTec's citation information will be updated upon publication.

@inproceedings{shao2020finegym,
  title={FineGym: A Hierarchical Video Dataset for Fine-grained Action Understanding},
  author={Shao, Dian and Zhao, Yue and Dai, Bo and Lin, Dahua},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages={2616--2625},
  year={2020}
}

License

This dataset is licensed under Creative Commons Attribution 4.0 International (CC-BY-4.0).
You are free to share and adapt the material, even commercially, as long as appropriate credit is given.

Note: The underlying video data remains the property of its original sources (e.g., YouTube). This dataset only distributes extracted pose annotations, not raw videos.

Acknowledgements

• Skeletons were extracted using pose estimators HRNet on the FineGym video corpus.
• We thank the authors of FineGym for their foundational work in fine-grained action recognition.