SVFSearch: A Multimodal Knowledge Intensive Benchmark for Short Video Frame Search in the Gaming Vertical Domain Paper: SVFSearch: A Multimodal Knowledge Intensive Benchmark for Short Video Frame Search in the Gaming Vertical Domain Project Page: https://svfsearch.github.io/SVFSearch page/ Code: https://github.com/SVFSearch/SVFSearch code Overview SVFSearch is the first open benchmark for short video frame search in the Chinese gaming vertical domain. Each example is constructed from a real short video clip centered on a paused game scene. The system must combine visual understanding of the paused frame with external game domain knowledge to answer a multiple choice question. SVFSearch evaluates a broad spectrum of systems — from direct QA multimodal LLMs to retrieval augmented workflows and fully autonomous multimodal agents — under a unified, reproducible offline retrieval environment with a frozen text corpus and image gallery. Key Features 9,198 high quality QA instances (5,000 test / 4,198 train) grounded in real short video frames 221 popular games spanning diverse genres, with 22,800 game specific core elements Frozen offline retrieval environment with a 45,608 entry text kn…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy