Leveraging Motion Estimation for Efficient Bayer-Domain Video Convolutional Networks
This work proposes Motion Estimation-based Video Convolution (MEVC), a novel framework that bypasses traditional image signal processors (ISP) by processing raw Bayer data directly. It integrates motion estimation into convolutional layers to reduce redundant calculations across temporal frames, achieving over 70% FLOPs reduction while maintaining high accuracy for video tasks.