紀要論文

動画像における行動認識での自己教師あり学習の有効性

EFFECTIVENESS OF SELF-SUPERVISED LEARNING FOR ACTION RECOGNITION IN VIDEOS

jpn

24368083

https://doi.org/10.15002/00030734

法政大学大学院理工学研究科

法政大学大学院紀要. 理工学研究科編

65

1

5

2024-03-24

Version of Record

In supervised learning, achieving high accuracy often requires a large labeled dataset; however, the associated human labeling costs can be high. The aim of this study is to improve recognition accuracy using a small labeled dataset in the context of action recognition in videos by employing self-supervised learning for feature extraction. In the pretraining phase, three-dimensional convolutional neural networks (3DCNN) are trained on frame order prediction to learn temporal features in video sequences. The pretrained parameters are used to initialize the 3D-CNN, which is then fine-tuned for action recognition as the main task. The action recognition is analyzed using training dataset sizes of 500, 2,500, 5,000, 10,000, 20,000, and 30,050 instances. For small training dataset sizes of 500, 2,500 and 5,000 instances, the pretrained model shows a significant improvement in recognition accuracy compared to the model without pretraining, confirming the effectiveness of self-supervised learning.

Article