ObjectStream framework uses latent objects for streaming video understanding · arXiv

Researchers have introduced ObjectStream, a novel framework designed to enhance streaming video understanding by using latent objects as memory anchors. This training-free approach directly extracts spatially coherent latent objects from frozen Video-LLM representations, linking them across frames to maintain persistent histories within a bounded memory budget. ObjectStream preserves object histories, changes, and recent visual context, enabling Video-LLMs to reason about object identities, interactions, and state changes without altering the base model. Experiments show significant improvements in performance and efficiency, with ObjectStream boosting Qwen2.5-VL-7B by 10.0 points on the OVO-Bench Real-Time Visual Perception benchmark while reducing memory usage and time-to-first-byte by approximately 50%. AI

IMPACT Enhances Video-LLM capabilities for real-time analysis and long-video comprehension, potentially improving applications in surveillance, content moderation, and automated video summarization.

RANK_REASON This is a research paper describing a new framework for video understanding. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• Influence Flower
• Litmaps
• ObjectStream
• OVO-Bench Real-Time Visual Perception
• Qwen2.5-VL-7B
• ScienceCast
• scite Smart Citations
• Video-LLM

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →