DaMO
收藏资源简介:
DaMO 是一个针对视频领域的大型语言模型,专门设计用于精确的时间推理和多模态理解。该模型的核心是 Temporal-aware Fuseformer,它采用分层双流架构,逐步捕捉每个模态中的时间动态,并有效地融合互补的视觉和音频信息。DaMO 还集成了一个全局残差,减少了空间冗余,同时保留了重要的语义细节。为了进一步增强计算效率,DaMO 还采用了一个分阶段的训练模式,逐步为模型配备多模态对齐、语义接地和时间推理能力。这项工作还通过 GPT 生成的具有时间定位的 QA 对,丰富了现有的数据集,为需要时间监督的任务提供了一个有价值的资产。
DaMO is a large language model tailored for the video domain, specifically designed for precise temporal reasoning and multimodal understanding. The core of this model is Temporal-aware Fuseformer, which adopts a hierarchical dual-stream architecture to gradually capture temporal dynamics within each modality and effectively fuse complementary visual and audio information. DaMO also integrates a global residual mechanism to reduce spatial redundancy while preserving important semantic details. To further enhance computational efficiency, DaMO adopts a staged training paradigm to gradually endow the model with multimodal alignment, semantic grounding and temporal reasoning capabilities. This work also enriches existing datasets with GPT-generated QA pairs with temporal localization, providing a valuable asset for tasks requiring temporal supervision.




