Answers are grounded in this paper and cite their sources.
As a result of the increasing popularity of the collection of multimodal data (motion capture videos), new possibilities for developing data-driven models of athletes’ motions appear. However, existing methods usually require single modality input or do not contain structures that could help to model the temporal evolution of movements and account for biomechanical aspects of motions and their development. To address those problems, we present an approach called Athletic on, which aims at developing a structured multimodal deep-learning-based solution for athlete motion modeling and performance representation. The method utilizes the use of biomechanics-constraint-aware encoding, relational attention modeling, and semantic-temporal decoding in order to develop representations of motion sequences based on heterogeneous input data. In addition, it includes a reasoning module (Stratiform), which is responsible for the refinement of motion trajectories by means of intention-aware correction and regularization of trajectory representation. The developed solution was tested using several benchmarks (Human3.6M, NTU RGB + D 120, Kinetics-700, and UCF101) for motion and action recognition and analysis. The comparison with existing approaches proved the superiority of the developed architecture in terms of accuracy, F1-score, and area under the curve. The proposed model is tested for its suitability as a multimodal transformer-based method for learning representations and action recognition tasks, but not as a personalized training/coaching tool.
No takes yet. Share an insight, caveat, or question.
Jiwang Sun (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: