Simple to Complex Transfer Learning for Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Human Action Recognition

Background:

• Recognizing complex human actions is challenging due to the difficulty in acquiring large amounts of labeled data.
• Complex actions can be decomposed into sequences of simpler actions, which are more readily available in existing datasets.

Purpose of the Study:

• To propose a novel Simple to Complex Action Transfer Learning Model (SCA-TLM) for enhanced complex human action recognition.
• To leverage abundant labeled simple actions to improve the performance of complex action recognition models.

Main Methods:

• The SCA-TLM optimizes weight parameters to reconstruct complex actions from simple actions.
• Optimal reconstruction coefficients are obtained by minimizing an objective function.
• Target model weights are represented as a combination of source model weights.

Main Results:

• The proposed SCA-TLM effectively utilizes simple actions for complex action recognition, outperforming methods relying solely on complex action samples.
• Extensive experiments on the Olympic Sports and UCF50 datasets demonstrate the model's effectiveness.
• The model shows significant improvements in complex action recognition performance.

Conclusions:

• The SCA-TLM offers a viable solution for complex human action recognition by effectively transferring knowledge from simple actions.
• This approach addresses the data scarcity issue in training robust action recognition models.
• The method provides a new direction for leveraging existing simple action datasets for complex action recognition tasks.