Related Experiment Video
Updated: Aug 19, 2025
08:25
Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment
Published on: May 7, 2019
9.1K
DANet: Semi-supervised differentiated auxiliaries guided network for video action recognition
Guangyu Gao1, Ziming Liu2, Guangjun Zhang1
1School of Computer Science and Technology, Beijing Institute of Technology, Beijing 100081, China.
Summary
This study introduces DANet, a semi-supervised network for Video Action Recognition (ViAR) that effectively uses limited annotated videos by leveraging knowledge from pre-trained image models. DANet significantly improves ViAR performance with minimal labeled data.
Area of Science:
- Computer Science
- Artificial Intelligence
- Machine Learning
Background:
- Deep Learning (DL) has advanced Video Action Recognition (ViAR), but performance relies heavily on large annotated datasets.
- Acquiring extensive video annotations is costly and time-consuming, hindering practical applications.
- Existing methods often suffer performance degradation with limited labeled data.
Purpose of the Study:
- To propose an end-to-end semi-supervised network, DANet, that maximizes the utility of limited annotated videos for ViAR.
- To leverage knowledge from pre-trained static image models to improve ViAR performance without requiring additional annotations.
- To develop a method that mitigates the performance gap caused by data scarcity in ViAR.
Main Methods:
- Introduced a Differentiated Auxiliary guided Network (DANet) with a two-branch architecture: a target ViAR branch and an auxiliary branch with multiple pre-trained models.
- Employed a semi-supervised learning strategy combining supervised cross-entropy loss on annotated videos and self-supervised contrastive losses on all videos (using auxiliary networks).
- Investigated different weighting strategies for auxiliary networks to guide the ViAR model, exploiting the connection between static image tasks and video recognition.
Main Results:
- DANet demonstrated superior performance on several action recognition benchmarks compared to state-of-the-art methods.
- The proposed method significantly outperformed existing ViAR techniques, especially when trained with very few annotated videos.
- Experimental results validated the effectiveness of the semi-supervised approach and the guidance from auxiliary networks.
Conclusions:
- DANet offers an effective solution for Video Action Recognition in low-data regimes.
- The approach successfully utilizes unlabeled data and knowledge from pre-trained image models to enhance ViAR.
- DANet represents a significant advancement in semi-supervised learning for action recognition, addressing practical data limitations.
