本研究提出了一种名为差异蒸馏的方法，旨在降低视频语言模型处理长视频的计算成本。通过差异关键帧选择和特征合并机制，开发了ViLaMP模型，能够在单个NVIDIA A100 GPU上高效处理最多10K帧的视频，并在多个视频理解基准测试中表现优异。

本研究提出了一种语义驱动的搜索框架，解决长视频理解中文本查询与视觉元素之间的逻辑关系。通过定义四种逻辑依赖关系，动态更新帧采样分布，实现语义关键帧识别，显著提升了关键帧选择和视频问答任务的性能。

本研究提出了一种频率自适应低延迟目标检测器（FAOD），有效解决了事件相机与RGB相机在目标检测中的数据融合问题，显著提高了检测精度，尤其在多个数据集上表现突出。

建立等变消息传递框架，实现几何深度学习中的任意维度欧几里得空间上的等变信息传递，通过局部坐标框架和张量对象在消息中保持几何信息一致性，并展示了 O (3) 等变性在点云架构上的优势及在法向量回归中的最新成果。

When I suggested a department chair in 2019 that they may consider opening a blog so that all students and faculty in the department can write together (which sounded exciting to me), he expressed...

In this blog we’re going to take a closer look at “connection coalescing”, with specific focus on manage it at a large scale

In this blog we’re going to take a closer look at “connection coalescing”, with specific focus on manage it at a large scale

In order to get enough information from videos, we often need to extract video frames. However, we do not want to extract every frame from a video due to: information redundancy in successive...

Hi everyone! Over the summers I did research on Wi-Fi and part of my research involved sending and receiving different kinds of IEEE 802.11 packets. I did most of this stuff using Scapy so I...