Aligning views improves evaluation and performance of 3D image synthesis
Aligning Misaligned Benchmarks for Single-View Novel View Synthesis
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Measuring how well computers create new images from a single photo has problems because the camera positions used aren’t always lined up correctly. The authors show that fixing how these camera views line up using simpler image matching improves how we judge these methods. When aligned properly, some 3D-based methods actually perform better than previously thought. They also present an easy way to generate new views using existing video tools without special training.
What this means in practice
- •For computer graphics developers: Improve how 3D image synthesis methods are evaluated and compared by aligning poses using image-based matching techniques.
- •For augmented reality app developers: Use the authors’ baseline that generates new views from a single image with off-the-shelf video tools and geometry estimators without extra training.
Authors
Yang-Che Tseng, Erqun Dong, Jeong-gi Kwak, Gopal Sharma, Shakiba Kheradmand, Stephan Garbin, Hossam Isack, Kwang Moo Yi
Abstract
While single-view novel view synthesis methods have been dramatically enhanced recently, how to quantify their performance has yet to be standardized. In this work, we demonstrate that an important factor lies in how each method's own reference frame is aligned with the so-called ground-truth pose. Existing methods often rely on test-time optimization of target poses for this alignment, which inherently is ill-posed due to the complexity of the optimization landscape of camera poses vs render accuracy, leading to methods being evaluated at an arbitrary local optimum. Some recognize this limitation and thus rely on 3D reconstructions of entire scenes using training and test views to align poses, which may be infeasible. In this work, instead, we propose to use traditional 2D-to-2D and 2D-to-3D image matching techniques to align 3D reconstructions. We also evaluate only within co-visible regions, between different views, which is the only place where a `correct' ground truth can be defined. Our evaluations show that with proper coordinate alignment, novel-view synthesis methods that use explicit 3D Gaussians achieve the state of the art, despite being previously reported in the literature to be outperformed by pure image-based 2D methods. We further provide a simple baseline that directly utilizes an off-the-shelf video generator and a feed-forward geometry estimator. Our simple baseline is already comparable to state-of-the-art methods specifically trained for novel-view synthesis, without any task-specific training.