Thmen/EGVSR

What it solves

This project addresses the challenge of Video Super-Resolution (VSR), which is the process of increasing the resolution of low-quality video sequences while maintaining visual quality and temporal coherence. Specifically, it aims to improve the inference speed of existing models like TecoGAN without sacrificing performance.

How it works

EGVSR (Efficient & Generic Video Super-Resolution) is a PyTorch implementation that utilizes subpixel convolution to optimize the inference speed of the TecoGAN VSR model. It provides a unified framework that allows users to implement and compare various deep learning-based VSR methods, including VESPCN, SOFVSR, FRVSR, and TecoGAN.

Who it’s for

This tool is designed for researchers and developers working in computer vision and video processing who need a high-performance framework for upscaling videos using state-of-the-art deep learning techniques.

Highlights

• Unified Framework: Supports multiple VSR methods (VESPCN, SOFVSR, FRVSR, TecoGAN, and EGVSR) in one repository.
• Optimized Speed: Uses subpixel convolution to achieve faster inference speeds compared to prior methods.
• Extensive Testing: Includes support for standard datasets like Vid4 and Tos3, as well as a new, more diverse dataset called Gvt72.
• Performance Metrics: Evaluates models based on objective visual quality and temporal coherence.