基于视频的人类行动识别的知识蒸:一种直观的方法,以高效和灵活的模型培训
Fernando Camarena1, Miguel Gonzalez-Mendoza1, Leonardo Chang2
1School of Engineering and Science, Tecnologico de Monterrey, Nuevo León 64700, Mexico.
Journal of imaging
|April 26, 2024
概括
知识蒸 (KD) 通过提高准确性和灵活性来增强自我监督的视频模型培训. 这种方法加速了融合,即使数据有限,为各种应用提供了可适应的解决方案.
科学领域:
- 计算机视觉 计算机视觉
- 机器学习 机器学习
- 人工智能的人工智能
背景情况:
- 在视频中训练人类行动识别模型是计算要求很高的.
- 当前的转移学习方法缺乏灵活性和效率,通常依赖于限制性的预训练架构.
研究的目的:
- 探索知识蒸 (KD) 以提高自我监督的视频模型培训.
- 为了提高分类准确性,加快融合,增加模型灵活性.
- 在定期和有限数据场景中评估KD的有效性.
主要方法:
- 应用知识蒸 (KD) 以指导自主监督视频模型的培训.
- 在UCF101数据集上测试了该方法,数据比例不同 (100%, 50%, 25%, 2%).
- 将KD指导的培训与传统的培训方法进行比较.
主要成果:
- 知识的蒸优于传统的培训,而不会影响分类的准确性.
- 在标准环境和数据稀缺环境中,KD降低了收率.
- 实现了跨架构的灵活性,用于各种应用,从资源有限到高性能.
结论:
- 知识蒸是提高自主监督视频模型培训效率和灵活性的一种有效技术.
- 对于数据有限的场景,KD提供了可行的解决方案.
- 该方法允许在不同的计算约束中进行可适应的模型定制.
相关概念视频
01:06Fixed Action Patterns
16.0K
A fixed action pattern (FAP) is a specific, hard-wired sequence of behaviors that occurs in response to an external stimulus, called a sign stimulus. The behavior is “fixed” because it is essentially unchangeable—proceeding similarly across individuals of a species every time it occurs.
16.0K
01:22Force Classification
1.2K
Forces play a crucial role in the study of physics and engineering. They are essential in describing the motion, behavior, and equilibrium of objects in the physical world. Forces can be classified based on their origin, type, and direction of action.
Contact and non-contact forces are two of the most widely used categories of forces. As the name suggests, contact forces require physical contact between two objects to act upon each other. Examples of contact forces include frictional,...
Contact and non-contact forces are two of the most widely used categories of forces. As the name suggests, contact forces require physical contact between two objects to act upon each other. Examples of contact forces include frictional,...
1.2K