An Evaluation of Quality Stability in NeRF Object Modeling

Neural Radiance Field (NeRF) has become an increasingly popular approach for modeling the geometry and appearance of scenes/objects. Various datasets for training and testing the rendered view quality have been introduced. However, there is still a lack in evaluating quality variations for NeRF models. This is especially crucial in interactive multimedia services like e-learning and Virtual Reality, where users may change the viewing direction frequently. To address this limitation, we carry out quality stability assessments with different NeRF models and 3D objects. Rather than using commonly used spiral sampling or random sampling on the target area, we employ parallel contour lines as the trajectory of cameras. The empirical results show that quality variations strongly depend on models, objects, and the local location (e.g. angle and height) of rendered images. Also, it is important to report statistic measures of quality variations in NeRF modeling.