DynamicZoom is a cutting-edge video enhancement system developed as part of the Computer Vision (CS/ECE 766) course at UW-Madison. The project addresses the common problem of video quality degradation during digital zoom by implementing real-time super-resolution technologies.

Key Features

  • Real-time video resolution enhancement
  • Region of Interest (ROI) based processing
  • GPU-accelerated processing using PyTorch and CUDA
  • Support for multiple upscaling models
  • Minimal processing latency (7-8ms for 720p to 4K)

Performance Highlights

  • Achieves high-quality upscaling with inference speeds of 7-8ms
  • Comparable quality to SwiftSRGAN with significantly faster processing
  • Implements optimized Bicubic++ architecture for real-time performance
  • Optimized for real-time applications in sports broadcasting, surveillance, and AR/VR

Applications

  • Sports Analysis: Enhanced slow-motion replays
  • Augmented Reality: Improved visual clarity
  • Security Systems: Better surveillance footage
  • Search and Rescue: Enhanced detail in critical operations

Technologies Used

  • Python
  • PyTorch
  • CUDA
  • Bicubic++ Model
  • Advanced Parallel Processing

Team:

  • Shantanu Vichare
  • Dario Placencio
  • Nico Ranabhat
  • Hemanth Sridhar Nakshatri