Core Concepts
The AIS 2024 Real-Time Image Super-Resolution (RTSR) Challenge introduced a novel benchmark for upscaling compressed 540p AVIF images to 4K resolution in real-time on commercial GPUs. The challenge attracted 160 participants, with 25 teams submitting their code and models. The solutions present novel designs tailored for memory-efficiency and runtime on edge devices, achieving significant PSNR improvements over Lanczos interpolation while processing images under 10ms.
Abstract
The AIS 2024 RTSR Challenge aimed to advance the state-of-the-art in real-time super-resolution (SR) of compressed images. The challenge used a diverse test set of 4K images ranging from digital art to gaming and photography, compressed using the modern AVIF codec.
The key highlights of the challenge and the submitted solutions are:
Motivation: The challenge leveraged AVIF as the image coding format to evaluate quality improvement from SR when combined with the superior compression efficiency of AVIF compared to JPEG.
Dataset: The 4K RTSR benchmark dataset included 110 diverse test samples, comprising real-world captures, rendered gaming content, and various digital art and photo-realistic images.
Evaluation: The participating teams provided their code, models and results, which were then validated and executed by the organizers to obtain the final results. The models were evaluated on PSNR, SSIM, runtime, and computational complexity (MACs).
Key Techniques: The top solutions utilized techniques like re-parameterization, pixel shuffle/unshuffle, multi-stage training, and knowledge distillation to achieve high efficiency and performance. Edge-oriented filters and global residual connections were common architectural choices.
Results: All the proposed methods improved PSNR fidelity over Lanczos interpolation and processed images under 10ms on an NVIDIA 4090 GPU. The best models achieved PSNR-Y improvements of up to 1.2dB over the baseline while maintaining real-time performance.
The challenge and the survey of the top solutions provide valuable insights into the state-of-the-art in efficient and real-time super-resolution of compressed high-resolution images.
Stats
The average PSNR-Y for QP31 and QP63 is 33.09 dB.
The average runtime across all methods is 2.72 ms.
The average MACs operations is 7.97 G.
Quotes
"All the proposed methods improve PSNR fidelity over Lanczos interpolation, and process images under 10ms."
"Considering the best methods, we can conclude that there is certain convergence in the model designs. As previously mentioned, re-parameterization is ubiquitous."