Ensemble Deep Learning and Algorithmic Techniques for Video Stabilization

Abstract

This paper explores various techniques for video stabilization, incorporating two algorithmic methods: Affine Transformation-Based Video Stabilization with Trajectory Smoothing and Frame Blending (ATTSFB), and Integrated Video Stabilization and Enhancement (IVSE). Additionally, a simple U-Net model and a simple Inception model were individually applied for stabilization. A tailored U-Net optical flow-based model was also employed for motion estimation and com-pensation. Finally, all three models were integrated using an ensemble approach employing the averaging technique. Despite these efforts, resource limitations constrained the stabilization results' effectiveness. The best-achieving model was the simple U-Net model of 0.5659 SSIM, and the least was the Inception model of 0.523 SSIM.