What Should a Streaming Video Model Remember?
Researchers have developed SelectStream, a novel framework for streaming video understanding models that selectively allocates memory to address the challenge of processing continuous video data under fixed computational budgets. Unlike previous methods that indiscriminately inject history, SelectStream prioritizes current scene perception by using a query-conditioned evidence budget. The framework employs surprise-driven windowing, priority-preserving consolidation, and graph reasoning to manage memory, achieving strong performance on benchmarks like StreamingBench and OVO-Bench. AI
IMPACT Improves efficiency and performance of AI models processing continuous video data.