Researchers from CERTH-ITI have detailed their approach for the TRECVID 2025 Ad-hoc Video Search and Video Question Answering tracks. For ad-hoc video search, they employed a two-stage framework using multiple vision-language models and a multimodal large language model (MLLM) for reranking to improve relevance scores. In the video question answering tasks, they fine-tuned an audio-visual MLLM and utilized frame subset ensembling and response aggregation techniques to achieve top scores. AI
IMPACT Presents novel multimodal LLM applications for video search and question answering, potentially advancing AI capabilities in content analysis.
RANK_REASON Paper detailing methods for a specific research benchmark (TRECVID 2025). [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
