Use AI to rate the game’s image quality! Intel launches new evaluation tool CGVQM

Intel recently opened the AI-based video quality evaluation tool CGVQM (Computer Graphics Visual Quality Indicators) on GitHub, aiming to provide objective quantitative standards for games and real-time rendering of images. The tool is published in PyTorch form, and the supporting research papers are released simultaneously.

　　Current games generally rely on super-scoring technologies such as DLSS to improve performance, but they are prone to cause visual problems such as ghosting and flickering. Traditional picture quality indicators (such as PSNR) only evaluate compression artifacts and cannot fully reflect the complex distortion of real-time rendering. To this end, the Intel team built a CGVQD dataset, covering diversified image quality degradation caused by technologies such as path tracing and neural hypersampling, and trained the CGVQM model based on this dataset. This model adopts a 3D convolutional neural network (3D-ResNet-18 architecture), which can simultaneously capture image features in space-time dimensions and more accurately identify dynamic image quality problems.

　　Experiments show that the CGVQM evaluation effect comprehensively exceeds existing tools: the complex version of CGVQM-5 is close to the baseline of human scoring, and the simplified version of CGVQM-2 is also firmly in the top three, and has shown good generalization ability in untrained videos. In the future, model performance can be further optimized by introducing Transformer or optical flow information.