Evaluation Framework for Novel View Synthesis

The generation of 3D views from 2D images has become an essential task in various fields such as 3D animation, computer vision and industrial digital twins. Over the past few years, novel view synthesis models, particularly neural-based approaches, have shown significant advantages over traditional photogrammetry techniques. Despite these advancements, evaluating and comparing different models remains challenging due to varying datasets, testing methodologies, and complex setup processes. To address these issues, we propose a flexible and scalable evaluation framework that facilitates the integration and comparison of both existing and novel 3D view synthesis models. This web-based framework simplifies dataset management and model evaluation through a user-friendly interface, supporting multiple input formats and state-of-the-art tools like COLMAP for dataset creation. It also integrates models such as Instant Neural Graphics Primitives, 3D Gaussian Splatting, and 4D Gaussian Splatting, enabling fair performance comparisons across different datasets and scenarios. These models are also evaluated regarding key aspects showing the usability and versatility of the framework.