ECCV 2026 | 长视频Token剪枝新范式：从关键帧到证据链

TL;DR - SemVID is a training-free visual-token pruning method for long-video temporal grounding that preserves an “evidence chain” rather than isolated keyframes. It reduces inference cost while maintaining accurate event-boundary localization under aggressive token compression.

• SemVID allocates per-frame token budgets using both query relevance and inter-frame changes, retaining evidence across the event timeline.
• It preserves complementary object, motion, and context tokens to capture relevant entities, temporal transitions, and scene continuity.
• Motion tokens serve as cross-frame relay nodes, while MMR selection prevents redundant object patches from consuming the token budget.
• On Charades-STA and ActivityNet-Grounding with Qwen3-VL and Qwen2.5-VL, SemVID outperformed existing pruning methods in localization and evidence-retention/connectivity metrics at equal budgets, particularly at low retention rates.