ProactiveFrame
A training-free inference-time strategy for streaming video VLMs. Through text tokens, the model actively adjusts the input frame sampling rate. Combined with a two-layer sliding window, it maintains a memory structure with high FPS for recent frames and sparse sampling for distant frames.
Timeline (last 90 days)
ProactiveFrame is a training-free baseline method that lets the model actively adjust the input frame rate through text tokens.
ProactiveFrame uses a two-level sliding window mechanism to manage historical information: recent frames retain high-FPS observations, while older frames are downsampled into sparse history.
ProactiveFrame improves by 5.4 and 1.5 percentage points respectively compared with sparse uniform sampling.
ProactiveFrame performs far worse than the oracle-guided focusing strategy, indicating that the model can hardly infer from the video stream itself when finer temporal perception is needed.
All Facts (4)
ProactiveFrame uses a two-level sliding window mechanism to manage historical information: recent frames retain high-FPS observations, while older frames are downsampled into sparse history.
50%UnverifiedProactiveFrame improves by 5.4 and 1.5 percentage points respectively compared with sparse uniform sampling.
50%UnverifiedProactiveFrame performs far worse than the oracle-guided focusing strategy, indicating that the model can hardly infer from the video stream itself when finer temporal perception is needed.
50%UnverifiedProactiveFrame is a training-free baseline method that lets the model actively adjust the input frame rate through text tokens.
50%