Paper Overview
- Field: Computer Vision
- Authors: Guangzhao He, Hadar Averbuch-Elor, Wei-Chiu Ma
- arXiv: 2609.20819
- Leverages 360° videos as omniscient ground truth, enabling reference-based evaluation even for objects outside the model's field of view.
- Proposes three evaluation aspects: 1. Object permanence — whether objects are remembered to still exist after leaving view 2. Motion continuity — whether object motion remains consistent over time 3. Appearance preservation — whether object appearance is faithfully maintained
- Paper: https://arxiv.org/abs/2609.20819
- Dataset and code: https://guangzhaohe.com/persistbench
Motivation
Perceiving and remembering the visual world is fundamental to navigating and interacting with our environment. Current 4D foundation models—such as camera-controllable video models or 4D reconstruction models—can perceive and reconstruct dynamic environments, but how well they remember what they have perceived remains an open question.
Existing benchmarks largely rely on pixel-level metrics and lack ground truth for objects once they leave the field of view, making them unable to evaluate visual memory in an object-centric manner against references.
PersistBench
To fill this gap, the authors introduce PersistBench, a dataset and metric suite that:
Findings
Evaluating a variety of models across multiple categories, the study finds that current models maintain only short-term consistency; performance degrades significantly once objects leave the field of view. The results highlight the gap between current model capabilities and robust visual memory—"seeing is not remembering"—providing guidance for the future development of 4D foundation models.