부산대학교 도서관

The 2nd International Knowledge Graph Reasoning Challenge involves social issues focusing on the safety of older adults in their homes. The challenge aims to extract statistical information related to actions and objects that pose risks to daily life. To answer each question in a video, we used Video-LLaVa, a large-scale visual language model (LVLM), using two approaches. The first approach involves inputting question text and video into Video-LLaVa. In this paper, we describe the results of zero-shot queries. The second approach is to obtain a detailed description of the video output using Video-LLaVa and then answer questions based on it. We have yet to achieve good results with these approaches, but we have identified some issues that we will discuss along with the results.