DeltaStreamer: Towards Human-Like Agentic Streaming Video Understanding

TL;DR

DeltaStreamer introduces an agentic framework for streaming video understanding that mimics human observation by tracking visual changes in structured memory and enabling targeted replay of past frames. It improves answer accuracy by retrieving missing visual evidence through query-driven planning and replay, outperforming prior methods on benchmark datasets.

Why It Matters

The approach enables more reliable and explainable video understanding in real-time applications where future queries are unknown during observation. By preserving dynamic event transitions and recovering lost details, it supports robust decision-making in surveillance, robotics, and interactive systems.

Key Findings

• DeltaStreamer maintains a change-aware memory that records visual transitions rather than static snapshots, preserving event dynamics.
• Query-driven planning identifies required evidence and triggers targeted replay of historical frames to recover omitted visual details.
• The framework achieves state-of-the-art performance on OVO-Bench (66.07%) and StreamingBench (59.53%), demonstrating superior accuracy and evidence grounding.