Fixing misaligned NVS benchmarks with image matching puts 3D Gaussians back on top
kwangmoo_yi · x · 2026-10-31
"Aligning Misaligned Benchmarks for Single-View Novel View Synthesis" (arXiv:2610.38349) shows single-view NVS evaluation lacks standardization, hinging on how each method's reference frame aligns with ground-truth poses.
- Common test-time pose optimization is ill-posed due to a complex landscape, evaluating methods at arbitrary local optima;
- Some methods align via full 3D scene reconstruction using train+test views, often infeasible;
- The paper instead uses traditional 2D-to-2D and 2D-to-3D image matching, and evaluates only within co-visible regions — the only place a 'correct' ground truth can be defined;
- With proper alignment, explicit 3D Gaussian methods achieve SOTA, reversing prior reports favoring pure image-based 2D methods; the authors also provide a simple baseline using an off-the-shelf video generator.
More from Research
- Aligning image and text embeddings without paired data: Wasserstein Procrustes and rotation papers — phillip_isola · 2026-10-31
- Phillip Isola demos aligning image and text embeddings with just a single rotation — phillip_isola · 2026-10-31
- Arena Offers Up to $50,000 Per Project for Academic Research on AI Evaluation — arena · 2026-10-31
- StudentBench: AI tutoring matches human GRE gains at 918x lower cost — gharik · 2026-10-31
- WISD: Distill Training Signal from Weaker Quantized Copies of Reasoning Models — jet-ai · 2026-10-31
- Albert Gu: H-Net learns to segment Chinese characters from raw bytes mid-training — niloofar_mire · 2026-10-31