KL-DNAS: Knowledge Distillation-Based Latency Aware-Differentiable Architecture Search for Video Motion Magnification

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Subtle motions, invisible to the naked eye, are crucial in various fields like biomechanics and material science.
• Current video motion magnification methods lack adaptability due to fixed computational complexity.
• Existing techniques struggle with applications requiring real-time processing or specific latency constraints.

Purpose of the Study:

• To develop an adaptable video motion magnification technique addressing limitations of state-of-the-art methods.
• To enable precise visualization of minute motions under varying time constraints.
• To improve the efficiency and accuracy of motion magnification for diverse applications.

Main Methods:

• Proposed a knowledge distillation-based latency-aware differentiable architecture search (KL-DNAS) for video motion magnification.
• Utilized a teacher network and knowledge distillation (KD) to reduce memory usage and enhance denoising.
• Incorporated search strategies for receptive fields and multi-feature connections, alongside a novel latency loss function.

Main Results:

• Achieved smaller model sizes compared to state-of-the-art methods.
• Demonstrated superior motion magnification performance with reduced visual distortions.
• Successfully optimized models for specific latency constraints while maintaining high output quality.

Conclusions:

• The KL-DNAS method offers a flexible and efficient solution for video motion magnification.
• This approach enhances the applicability of motion magnification in real-time and resource-constrained scenarios.
• The findings pave the way for improved analysis of subtle motions in scientific and medical applications.