DeltaStreamer
A streaming video understanding framework that mimics how humans watch while noting changes and verify only when asked. It combines change-aware memory, query-driven planning, and targeted replay to address ambiguous event transitions and unrecoverable missing evidence in streaming scenarios.
Timeline (last 90 days)
DeltaStreamer proposes a change-aware replay mechanism that stores only state deltas rather than full states, thereby greatly reducing memory usage and latency while preserving the temporal dynamics of video world models.
DeltaStreamer consists of three core components: change-aware memory, a query-driven planner, and targeted evidence collection and replay.
Change-aware memory explicitly writes visual changes into structured memory before a query arrives, and the memory entries are the 'changes' themselves rather than static snapshots.
After receiving a query, the query-driven planner determines what evidence is needed, whether the evidence is already in memory, and if not, which historical frames need to be revisited, outputting an evidence acquisition path rather than an answer.
The targeted replay mechanism allows the agent to revisit original video frames during the inference stage, recovering at the pixel level visual details that were not written into the text descriptions during the observation stage.
DeltaStreamer is a streaming video understanding method whose core design is to imitate the human approach of 'continuous observation combined with active investigation', rather than compressing video history into a single summary.
DeltaStreamer achieves an overall score of 59.53% on the StreamingBench benchmark.
DeltaStreamer achieves improvements over representative methods on both the OVO-Bench and StreamingBench benchmarks, and the paper emphasizes that the improvement comes from connecting continuous observation with active investigation rather than from a single module improvement.
DeltaStreamer's project homepage has been made public.
DeltaStreamer achieves an overall score of 66.07% on the OVO-Bench benchmark.
All Facts (10)
DeltaStreamer proposes a change-aware replay mechanism that stores only state deltas rather than full states, thereby greatly reducing memory usage and latency while preserving the temporal dynamics of video world models.
50%UnverifiedDeltaStreamer consists of three core components: change-aware memory, a query-driven planner, and targeted evidence collection and replay.
50%UnverifiedChange-aware memory explicitly writes visual changes into structured memory before a query arrives, and the memory entries are the 'changes' themselves rather than static snapshots.
50%UnverifiedAfter receiving a query, the query-driven planner determines what evidence is needed, whether the evidence is already in memory, and if not, which historical frames need to be revisited, outputting an evidence acquisition path rather than an answer.
50%UnverifiedThe targeted replay mechanism allows the agent to revisit original video frames during the inference stage, recovering at the pixel level visual details that were not written into the text descriptions during the observation stage.
50%UnverifiedDeltaStreamer is a streaming video understanding method whose core design is to imitate the human approach of 'continuous observation combined with active investigation', rather than compressing video history into a single summary.
50%UnverifiedDeltaStreamer achieves an overall score of 59.53% on the StreamingBench benchmark.
50%UnverifiedDeltaStreamer achieves improvements over representative methods on both the OVO-Bench and StreamingBench benchmarks, and the paper emphasizes that the improvement comes from connecting continuous observation with active investigation rather than from a single module improvement.
50%UnverifiedDeltaStreamer achieves an overall score of 66.07% on the OVO-Bench benchmark.
50%UnverifiedDeltaStreamer's project homepage has been made public.
50%