Idea
A fast, storage-efficient in-loop filtering platform for video codecs that improves visual quality and reduces bitrate for streaming providers and device makers.
Research Paper
Core Innovation
This paper presents LUT-ILF++, which replaces computationally expensive neural network inference in in-loop filtering with learned look-up tables and interpolation. It innovates by enabling multiple LUT cooperation, customized and cross-component indexing, and LUT compaction to reduce storage needs. This approach achieves significant bitrate savings with much lower complexity compared to prior DNN-based ILF methods.
Market Size (TAM)
$10–20B TAM, $2–5B SAM; assumption: growing demand for efficient video compression in streaming, gaming, and broadcasting sectors.
Potential Customers & Pain Points
- Video Streaming Platforms Needing Lower Bandwidth Costs
- Device Manufacturers Requiring Efficient Video Decoding
- Video Codec Developers Seeking Improved Compression
- Cloud Gaming Services Demanding Low Latency and High Quality
Business Model
Licensing the LUT-ILF++ technology to codec developers and streaming platforms; offering integration support and custom optimization services.
Competitive Landscape
- Google VVC Solutions
- Fraunhofer IIS
- Tencent Video Codec Team
Implementation Challenges
- Integration with existing codec standards
- Balancing LUT size and compression efficiency
- Adoption by industry stakeholders
Validation Strategy
- Implement LUT-ILF++ in popular VVC codec software and benchmark bitrate savings
- Conduct real-world streaming tests to measure quality and latency improvements
- Partner with device manufacturers for pilot deployments and feedback
Research Paper Overview
In-Loop Filtering Using Learned Look-Up Tables for Video Coding
Summary
In-loop filtering (ILF) reduces artifacts and enhances visual quality in video coding. Neural network-based ILF improves coding gains but is computationally expensive. This paper proposes LUT-ILF++, a practical ILF solution using learned look-up tables derived from a trained DNN to replace heavy inference with fast table lookups and interpolation. It introduces multiple LUT cooperation, customized indexing, cross-component indexing, and LUT compaction for storage efficiency. Implemented in VVC software, it achieves significant bitrate reductions with much lower complexity and storage than DNN-based methods.