Perceptual Quality Assessment of 3D Gaussian Splatting: A Subjective Dataset and Prediction Metric

The recent publication on 3D Gaussian Splatting (3DGS) highlights the creation of 3DGS-QA, a pioneering subjective quality assessment dataset comprising 225 degraded reconstructions across 15 object types. This initiative aims to fill the gap in understanding the perceptual quality of 3DGS-rendered content, which has been largely overlooked in previous research. Factors such as viewpoint sparsity, limited training iterations, point downsampling, noise, and color distortions can significantly impact visual fidelity, yet their perceptual effects have not been systematically studied until now. The introduction of a no-reference quality prediction model that operates on native 3D Gaussian primitives allows for the estimation of perceived quality without the need for rendered images or ground-truth references. This model has been benchmarked against existing quality assessment methods, demonstrating superior performance in evaluating the visual quality of 3DGS content. The findings from thi…