MA-LMM: Memory-Augmented Large Multimodal Model for Long-Term Video Understanding

With the success of large language models (LLMs), integrating the visionmodel into LLMs to build vision-language foundation models has gained much moreinterest recently. However, existing LLM-based large multimodal models (e.g.,Video-LLaMA, VideoChat) can only take in a limited number of frames for shortvideo understanding. In this study, we mainly focus on designing an efficientand effective model for long-term video understanding. Instead of trying toprocess more frames simultaneously like most existing work, we propose toprocess videos in an online manner and store past video information in a memorybank. This allows our model to reference historical video content for long-termanalysis without exceeding LLMs' context length constraints or GPU memorylimits. Our memory bank can be seamlessly integrated into current multimodalLLMs in an off-the-shelf manner. We conduct extensive experiments on variousvideo understanding tasks, such as long-video understanding, video questionanswering, and video captioning, and our model can achieve state-of-the-artperformances across multiple datasets. Code available athttps://boheumd.github.io/MA-LMM/.