When and Where to Look: Adaptive Visual Evidence Scheduling for Efficient Long Video Understanding

In 30 Sekunden

Was passiert ist

EcoFrame framework adaptively selects video frames for vision-language models using entropy-based budget scheduling and attention-guided search instead of static frame selection.

Warum es zählt

Matters for engineers building long-video understanding systems who need faster inference without sacrificing accuracy on benchmarks like Video-MME and LongVideoBench.

Achtung

Training-free approach requires the VLM to already be deployed; unclear how well entropy signals generalize across different model architectures and video domains.