Perceptual quality metric and loss function for 3D and temporal consistency

To better train and evaluate 3D reconstruction methods (NeRF, Gaussian Splatting) or 3D generative models, both for static (3D) and dynamic (4D) scenes, we will develop a new full-reference quality metric and no-reference loss function. Those will be trained and validated on a new 4D quality dataset, with the subjective quality measured in stereoscopic presentation (e.g., on a VR headset). The developed techniques will improve 3D and temporal consistency of the rendered views, resulting in fewer temporal artefacts. They will also allow automatic hyper-parameter tuning and more reliable evaluation and comparison of 3D rendering techniques.