GymPose: a new benchmark and pose-based framework for fine-grained action classification

Fine-grained action recognition in artistic gymnastics requires accurate modeling of subtle pose variations, temporal coordination, and execution phases that are not sufficiently captured by coarse action benchmarks. This study introduces GymPose, a pose-centric benchmark dataset for fine-grained gymnastic action classification. The dataset was collected in a real training environment using a synchronized six-camera Intel RealSense RGB-D setup and includes 16,782 annotated frames from 13 athletes. It covers six fundamental gymnastic actions and 18 phase-level micro-action classes obtained by dividing each action into three execution phases. Human pose keypoints were extracted using ViTPose and RTMPose, producing 25- and 26-keypoint representations, respectively, after YOLO11x-based person detection. The discriminative capacity of these pose representations was evaluated using DenseNet121, ResNeXt50, FT-Transformer, and XGBoost under a stratified 70/15/15 train-validation-test protocol. On the GymPose-18 task, DenseNet121 achieved 90.04% accuracy using ViTPose features, while FT-Transformer and XGBoost achieved 89.07% and 88.13%, respectively. Additional experiments on the Yoga-82 dataset further demonstrate the effectiveness and generalizability of pose-based representations for capturing subtle motion differences. The results indicate that GymPose provides a structured benchmark for phase-level micro-action analysis in gymnastics and can support future studies on pose-based sports analytics, multi-view modeling, and action quality assessment. The dataset and source code are publicly available at https://github.com/meyurtsever/gympose .