Paper Overview
- Research area: Computer Vision (CV)
- Authors: Jiaxin Wu, Yihao Pi, Yinling Zhang
- Published: 2026-05-16
- arXiv: 2505.08633
- Given a generated clip, object-centric observations are obtained via segmentation and point tracking (e.g., SAM 2, MegaSaM, and CoTracker3).
- Observations are lifted to 3D world-space coordinates via monocular reconstruction.
- A set of projective-geometry residuals is computed, capturing three failure dimensions: 1. Scale-depth alignment 2. 3D motion consistency 3. 3D structural rigidity
- Project page, code, and dataset: https://pdi-bench.github.io/
- arXiv: https://arxiv.org/abs/2505.08633
Summary
Generative video models are increasingly studied as implicit world models, yet evaluating whether they produce physically plausible 3D structure and motion remains challenging. Most existing video evaluation pipelines rely heavily on human judgment or learned graders, which can be subjective and weakly diagnostic for geometric failures.
The authors introduce PDI-Bench (Perspective Distortion Index), a quantitative framework for auditing geometric coherence in generated videos:
Dataset and Findings
To support systematic evaluation, the authors construct the PDI dataset, covering diverse scenes designed to test these geometric constraints. Across state-of-the-art video generators, PDI reveals consistent geometry-specific failure modes not captured by common perceptual metrics, offering a diagnostic signal for progress toward physically grounded video generation and physical world models.