TEMPURA: Temporal Event Masked Prediction and Understanding for Reasoning in Action

AI Digest

TEMPURA提出两阶段框架，通过事件掩码预测和视频分割提升视频时间理解，结合事件推理与细粒度时间分割，显著优于现有模型。

TEMPURA introduces a two-stage framework combining event prediction and segmentation to enhance video temporal understanding, outperforming existing models in benchmarks.

Key points

• 提出两阶段训练框架，结合事件掩码预测与视频分割 Proposes a two-stage framework integrating event masking prediction and video segmentation
• 基于VER数据集（50万视频时序标注）进行训练 Trained on VER dataset with 500K temporally annotated videos
• 在时间定位和高亮检测任务中超越基线模型 Outperforms baseline models in temporal grounding tasks
• 通过事件推理与细粒度时间分割提升理解效果 Enhances understanding via event reasoning and fine-grained segmentation

Takeaway: 结合事件推理与细粒度时间分割是提升视频理解的关键。 / Combining event reasoning with fine-grained temporal segmentation is key to improving video understanding.

Why it matters 提供创新框架，解决视频理解中的时间定位与事件推理难题，具有实际应用价值。