Online Memory for Dynamic View Synthesis
A method for online dynamic novel view synthesis that updates expensive memory periodically while applying memory every frame.
TL;DR — The abstract proposes an online novel-view-synthesis method for multi-view streaming videos. It separates expensive memory updates from memory use: memory is updated periodically but applied every frame. Cross-view attention handles deformation between stored memory and the current frame, while Memory Loss and Memory Caching aim to preserve historical scene information and reduce catastrophic drift. The abstract claims real-time, state-of-the-art performance and minute-scale online memorization, but reports no exact metrics, datasets, hardware, or citation metadata.
Core problem
The abstract addresses online novel view synthesis from multi-view streaming videos. Its central problem is a trade-off: the system needs persistent, long-horizon memory to reconstruct temporarily occluded regions, but it must also satisfy strict real-time constraints.
The dynamic setting matters because standard Test-Time Training models are described as requiring gradient-based memory updates at every frame to adapt to changing motion. According to the abstract, those heavy updates are too expensive for real-time application and can become unstable over long contexts.
Full analysis, extracted claims, numerical results, entity graph, FAQ, related work, applications, and BibTeX are available via x402 micropayment.