Adaptive convolutional network and transfer learning-based dance movement recognition for realistic subjects

As deep learning technology develops rapidly, dance action recognition, a crucial computer vision research direction, has been progressively applied in other disciplines. But classical action recognition techniques often fail in complicated contexts with various action categories and dynamically shifting backgrounds; dance action recognition for realistic topics faces many difficulties. This work thus suggests a dance action recognition model ACN-TL-DAR based on adaptive convolutional network and transfer learning (TL), which combines adaptive convolutional networks and TL to efficiently manage complicated dance action data. This work confirms the great performance of the ACN-TL-DAR model on several criteria by means of experimental evaluation on two datasets. The experimental results reveal that the model suggested in this work has strong robustness and efficient identification capacity in several contexts, thereby offering a fresh concept for the expansion of the field of realistic dance movement recognition.