The paper introduces PersistBench, a new benchmark that uses 360-degree videos to provide ground truth for objects that leave the camera's field of view. It evaluates three aspects of visual memory: whether objects continue to exist (object permanence), whether their motion stays consistent (motion continuity), and whether their appearance remains stable (appearance preservation). Testing multiple 4D models shows that while they maintain short-term consistency while objects are visible, performance degrades significantly once objects are no longer observed. This work reveals that current models can perceive dynamic scenes but lack persistent memory, highlighting a critical gap between seeing and remembering for embodied AI and robotics.
4D foundation models lack robust visual memory for out-of-view objects
New PersistBench benchmark uses 360° videos to show that current video and reconstruction models degrade sharply once objects leave the field of view.
Academic
Guangzhao He · Hadar Averbuch-Elor · Wei-Chiu Ma
Cornell University
Research Digest··2 min read
In one line
Current 4D foundation models fail to remember objects that exit their field of view.
Before you read this
An idea this paper builds onRead it as
The authors introduce PersistBench, a dataset and metric suite using 360° videos as ground truth to test object permanence, motion continuity, and appearance preservation in 4D foundation models.
The authors developed PersistBench, a benchmark comprising 360° videos that provide omniscient ground truth beyond the traditional field-of-view.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.