Related Experiment Video
Updated: Jul 6, 2026
10:02
FIM Imaging and FIMtrack: Two New Tools Allowing High-throughput and Cost Effective Locomotion Analysis
Published on: December 24, 2014
12.2K
A CNN-Bi-LSTM pipeline and open FSW dataset for freestyle wrestling action recognition
Milad Rostamian1, Ali Mottaghi2, Mohsen Soryani3
1School of Computer Engineering, Iran University of Science and Technology (IUST), Tehran, Iran.
Scientific Reports
|March 24, 2026
Summary
We developed a foreground-aware system for recognizing actions in freestyle wrestling, achieving 82.9% accuracy. This method improves human action recognition in close-contact sports despite occlusion challenges.
Area of Science:
- Computer Vision
- Sports Analytics
- Machine Learning
Background:
- Human action recognition in close-contact sports faces challenges like occlusion and rapid movements.
- Freestyle wrestling presents a complex scenario due to sustained physical interaction.
Purpose of the Study:
- To develop and evaluate a novel system for human action recognition in freestyle wrestling.
- To introduce the Open FSW dataset for studying wrestling techniques.
Main Methods:
- A foreground-aware RGB pipeline using DeepLabV3+ for segmentation.
- Feature extraction with CNN backbones (VGG16, InceptionV3, EfficientNet-B7).
- Aggregation of features using a bidirectional LSTM for clip-level predictions.
Main Results:
- The best configuration achieved 82.9% top-1 accuracy.
- Foregrounding provided consistent gains, especially for high-occlusion techniques.
- Ablation studies quantified the benefits of foreground segmentation.
Conclusions:
- The proposed method effectively recognizes actions in freestyle wrestling.
- The Open FSW dataset and code are released to facilitate further research.
- Findings are domain-specific to freestyle wrestling due to dataset size limitations.
