Aligning misaligned benchmarks for single-view novel view synthesis

Synthesizing novel views of a scene from a single image has advanced quickly in recent years, yet the field has no standardized way to measure how well these methods perform. This thesis identifies an overlooked factor at the heart of this problem: the alignment between each method’s own reference frame and the so-called ground-truth camera pose. Evaluations that align the two through test-time optimization of the target pose are ill-posed: the landscape relating camera pose to render accuracy is highly non-convex, so methods are scored at an arbitrary local optimum. Alternatives that align poses by reconstructing entire scenes in 3D from both training and test views are often infeasible. This thesis instead proposes an evaluation protocol that aligns 3D reconstructions using traditional 2D-to-2D and 2D-to-3D image-matching techniques, and that computes metrics only within regions co-visible between views, the only regions for which a ‘correct’ ground truth can be defined. Under this protocol, the reported rankings of existing methods change substantially. With proper coordinate alignment, methods built on explicit 3D Gaussian representations reach the state of the art, though earlier literature reports them to be outperformed by purely image-based 2D methods. Building on this observation, this thesis further presents a simple baseline that directly combines an off-the-shelf video generator with a feed-forward geometry estimator. Without any task-specific training, this baseline is already comparable to the state of the art.