Abstract

The use of video in our everyday digital interactions is on the rise. With the advancement of higher-resolution content analysis and displays, the substantial volume of video content presents considerable challenges in terms of acquisition, trans- mission, compression, and display while maintaining high quality. Video compression is very crucial to manage the resources as well as ensure smooth transmission and playback of digital videos across different platforms and devices. In our work, we introduce an optimized PWC-Net (Pyramidal Warping Cost Volume-Net) architecture for motion estimation, which is a computationally intensive initial component present in codecs for video compression, such as H.264 and H.265/HEVC. Proposed approach utilizes a CNN (Convolutional Neural Network) based PWC-Net architecture to estimate the optical flow between the frame sequences, which is considered as the motion information. Our model, compresses the video sequences efficiently without computing the motion vectors unlike the traditional motion es- timation techniques using block-based methods. It demonstrates its effectiveness with the experimental results when performing compression on various H.264 codecs. It learns efficient video compression without the need for motion computation unlike the traditional motion estimation techniques and our experiments show that an optimized PWC-Net outperforms good compression ratio by performing motion estimation through optical flow when compared with existing motion estimation schemes on H.264 and H.265 codec. Index Terms—