Description
EvTexture and its journal extension EvTexture++ represent advanced solutions for video super-resolution (VSR) by incorporating event-driven texture enhancement. Developed by researchers from the University of Science and Technology of China, these projects utilize data from event cameras, which capture motion information with high temporal resolution and low latency, to significantly improve the quality of upscaled videos.
EvTexture, presented at ICML 2024, focuses on enhancing texture details in videos during the super-resolution process. The subsequent journal extension, EvTexture++, accepted by IEEE TPAMI in 2026, further refines these techniques for even greater fidelity and robustness. Both implementations are provided as official PyTorch code, allowing researchers and developers to integrate and build upon this cutting-edge technology.
The repository offers comprehensive resources, including pre-trained models for various datasets and degradation types, such as REDS and Vimeo-90K, supporting 4x super-resolution. It also provides access to prepared test datasets for benchmarks like Vid4 and REDS4, which include both video frames and corresponding event data. Installation instructions are detailed for both Conda environments and Docker, ensuring accessibility for a wide range of users.
Key capabilities include the ability to process video data alongside event streams to reconstruct finer textures and reduce artifacts often seen in traditional VSR methods. The project is built upon the BasicSR framework and draws inspiration from other notable works like RAFT, event_utils, and EvTexture-jupyter. This foundation allows for efficient experimentation and deployment. The project is released under the Apache-2.0 license, encouraging open collaboration and further research in the field of event-driven video enhancement.
EvTexture & EvTexture++'s Core Features
Event-driven texture enhancement for video super-resolution
Official PyTorch implementation for EvTexture (ICML 2024)
Official PyTorch implementation for EvTexture++ (TPAMI 2026)
Includes pre-trained models for REDS and Vimeo-90K datasets
Supports 4x super-resolution scale
Provides preprocessed test datasets for Vid4 and REDS4
Includes video demos showcasing upsampling results
Offers installation via Conda and Docker
Built upon the BasicSR framework
Apache-2.0 license for open use and modification
Getting Started with EvTexture & EvTexture++
Clone Repository: Obtain the EvTexture code from the GitHub repository.
Install Dependencies: Set up the required Python environment with specified PyTorch and CUDA versions, or use the provided Docker image.
Download Models and Data: Acquire pre-trained models and preprocessed test datasets for Vid4 and REDS4.
Configure Testing: Adjust the test configuration files (e.g., options/test/EvTexture/test_EvTexture_Vid4_BIx4.yml).
Execute Testing: Run the distributed testing script with the specified number of GPUs.
Analyze Results: Review the generated inference results in the 'results/' directory.
EvTexture & EvTexture++'s Use Cases
- Video Super-Resolution
- Texture Enhancement
- Event Camera Data Processing
- Research in Video Restoration
- Real-time Video Upscaling
- Low-light Video Improvement
