The paper "DGL: Dynamic Global-Local Prompt Tuning for Text-Video Retrieval" addresses the challenge of text-video retrieval (TVR) by proposing a novel method called DGL (Dynamic Global-Local Prompt Tuning). The main issues with existing methods, such as CLIP4Clip and VoP, are the inability to capture global video information and the lack of effective cross-modal alignment. DGL aims to solve these problems by generating dynamic local-level prompts (text and frame prompts) from a shared latent space, ensuring cross-modal interaction and alignment. Additionally, DGL introduces a global-local video attention mechanism to capture both global and local video information, enhancing the model's ability to understand temporal dynamics and holistic video content. Key contributions of DGL include: 1. **Dynamic Cross-Modal Prompts**: DGL generates dynamic local-level prompts from a shared latent space, ensuring effective cross-modal interaction. 2. **Global-Local Video Attention**: This mechanism captures both global and local video information, improving the model's understanding of temporal dynamics. 3. **Efficient Parameter Tuning**: DGL achieves superior or comparable performance to fully fine-tuned methods while tuning only a fraction of the parameters. Experiments on four datasets (MSR-VTT, VATEX, LSMDC, and ActivityNet) demonstrate that DGL outperforms or matches the performance of fully fine-tuned methods with significantly fewer trainable parameters. The paper also includes a detailed analysis of the model's effectiveness through ablation studies and visualizations, highlighting the importance of global and local information in text-video retrieval tasks.The paper "DGL: Dynamic Global-Local Prompt Tuning for Text-Video Retrieval" addresses the challenge of text-video retrieval (TVR) by proposing a novel method called DGL (Dynamic Global-Local Prompt Tuning). The main issues with existing methods, such as CLIP4Clip and VoP, are the inability to capture global video information and the lack of effective cross-modal alignment. DGL aims to solve these problems by generating dynamic local-level prompts (text and frame prompts) from a shared latent space, ensuring cross-modal interaction and alignment. Additionally, DGL introduces a global-local video attention mechanism to capture both global and local video information, enhancing the model's ability to understand temporal dynamics and holistic video content. Key contributions of DGL include: 1. **Dynamic Cross-Modal Prompts**: DGL generates dynamic local-level prompts from a shared latent space, ensuring effective cross-modal interaction. 2. **Global-Local Video Attention**: This mechanism captures both global and local video information, improving the model's understanding of temporal dynamics. 3. **Efficient Parameter Tuning**: DGL achieves superior or comparable performance to fully fine-tuned methods while tuning only a fraction of the parameters. Experiments on four datasets (MSR-VTT, VATEX, LSMDC, and ActivityNet) demonstrate that DGL outperforms or matches the performance of fully fine-tuned methods with significantly fewer trainable parameters. The paper also includes a detailed analysis of the model's effectiveness through ablation studies and visualizations, highlighting the importance of global and local information in text-video retrieval tasks.